4h-sic wafer intelligent laser slicing method and system

Abstract

The application belongs to the technical field of cross technology of semiconductor material processing, ultrafast laser internal modification and artificial intelligence process optimization, and proposes a 4H-SiC wafer intelligent laser slicing method and system, obtains the resistivity, doping type and target focal depth of the wafer to be sliced, generates continuous main stem input features and discrete condition vectors respectively, outputs a laser slicing process parameter set through the main stem network combined with the channel-by-channel affine modulation reasoning, generates a spherical aberration compensation phase diagram according to the target focal depth to obtain a corrected beam, completes internal modification scanning of the wafer by using the corrected beam and applies stress to realize wafer peeling, collects the wafer quality index after peeling, and expands the process database or triggers neural network incremental learning. The application can adapt to the processing conditions of different doped wafers and improve the adaptation effect of 4H-SiC wafer laser slicing processing.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of semiconductor material processing, ultrafast laser internal modification, and artificial intelligence process optimization, specifically to a method and system for intelligent laser slicing of 4H-SiC wafers. Background Technology

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] 4H-SiC, as a core substrate material for third-generation wide-bandgap semiconductors, is widely used in the fabrication of new energy vehicles, high-voltage power transmission and distribution, and radio frequency devices. Semiconductor device performance places high demands on the processing quality of substrate wafers, making wafer slicing a critical step in the substrate fabrication process. Traditional diamond wire saw slicing methods are prone to kerf loss and internal substrate damage. Ultrafast laser internal modification slicing, relying on its non-contact processing characteristics, is gradually becoming an alternative processing method for silicon carbide substrate slicing. This processing method focuses a laser into the interior of the crystal to create a modified region, and then uses stress to complete wafer separation. It can adapt to the thinning requirements of wide-bandgap semiconductor substrates and is gradually being promoted and applied in semiconductor substrate processing scenarios.

[0004] Different doping types of 4H-SiC wafers exhibit varying photoresponse characteristics. Current ultrafast laser slicing processes typically rely on fixed process parameters or repeated manual adjustments to determine processing conditions. When dealing with batches of wafers with differing doping properties, existing methods struggle to comprehensively match the entire laser processing setup with the wafer's electrical properties and the target processing conditions. Furthermore, beam focusing distortion caused by the dielectric material during processing is easily overlooked, and the slicing quality information after processing is difficult to use to optimize subsequent processing decisions, potentially leading to poor processing adaptation. Summary of the Invention

[0005] To address the problem of poor processing adaptability of 4H-SiC wafers with different doping types being difficult to match with laser processing conditions, this invention provides a smart laser slicing method and system for 4H-SiC wafers. By acquiring wafer electrical and processing target information, corresponding feature vectors are generated. A set of laser slicing process parameters is obtained through conditional neural network inference. Combined with spherical aberration correction, a corrected beam is obtained for modified scanning. Wafer stripping is then completed, and process iteration is achieved using slicing quality information. This enables adaptive output of processing conditions for wafers with different doping types, improving the processing adaptability of 4H-SiC wafer laser slicing.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for intelligent laser slicing of 4H-SiC wafers.

[0007] A smart laser slicing method for 4H-SiC wafers includes the following steps: Obtain the resistivity, doping type, and target focal depth of the 4H-SiC wafer to be sliced; Logarithmic transformation and standardization are performed on the resistivity to generate continuous backbone input features. The doping type and the target focal depth are encoded using one-hot encoding and then concatenated to generate a discrete conditional vector. The continuous backbone input features are input into the backbone network to obtain the backbone hidden features. The discrete conditional vector is used to generate scaling factors and offset factors. The scaling factors and offset factors are used to perform channel-wise affine modulation on the backbone hidden features to obtain modulated hidden features. The set of laser slicing process parameters is output based on the modulated hidden features. Based on the target depth of focus, the focal shift is calculated to generate a spherical aberration compensation phase diagram. Based on the spherical aberration compensation phase diagram, the beam focusing distortion is corrected to obtain the corrected beam. Based on the laser slicing process parameter set, the corrected beam is used to perform internal wafer modification scanning, and stress is applied to the modified wafer to achieve wafer peeling. Collect wafer quality indicators after stripping, and expand the process database or trigger incremental learning of neural networks based on the wafer quality indicators after stripping.

[0008] In one implementation of the first aspect of the present invention, performing a logarithmic transformation and standardization on the resistivity to generate continuous backbone input features includes: The logarithmic resistivity is obtained by taking the common logarithm of the resistivity. Retrieve the mean log resistivity and standard deviation log resistivity pre-obtained from a large number of wafer samples within the process database; Subtract the mean of the logarithmic resistivity from the logarithmic resistivity, and then divide by the standard deviation of the logarithmic resistivity to complete the standardized calculation of zero mean unit variance. The dimensionless continuous backbone input features are obtained through the aforementioned standardized calculations. These continuous backbone input features are used to input the backbone network for subsequent feature extraction operations.

[0009] In one implementation of the first aspect of the present invention, generating scaling factors and offset factors using the discrete condition vector includes: The discrete conditional vector is completely input into the conditional modulation branch network; Two sets of independent linear mapping operations are performed within the conditional modulation branch network. The scaling factor is output through one of the sets of linear mapping operations; The offset factor is output through another set of linear mapping operations; The vector dimensions of the scaling factor and the offset factor obtained from the output are consistent with the dimensions of the hidden features of the backbone network output.

[0010] In one implementation of the first aspect of the present invention, the trunk hidden features are obtained by performing channel-by-channel affine modulation using the scaling factor and the offset factor, including: The scaling factor is multiplied element-wise with the trunk hidden feature according to the channel dimension; Obtain the intermediate result of the element-wise multiplication operation; The intermediate calculation result is then added element by element to the offset factor. After performing element-by-element addition, the modulated hidden feature is output. The modulated hidden features incorporate doping and depth-of-focus constraint information from the discrete condition vector, which are used for subsequent inference calculations to obtain the laser slicing process parameter set.

[0011] In one implementation of the first aspect of the present invention, the wafer internal modification scan is performed using the corrected beam according to the laser slicing process parameter set, including: Real-time acquisition of wafer surface height undulation data; The wafer surface height fluctuation data is converted into a control signal, and the axial height of the objective lens is adjusted in a closed loop to maintain the stability of the actual depth of the laser focus inside the wafer. The laser slicing process parameter set was analyzed to extract the laser power, scanning speed, scanning step size, and pulse repetition frequency. Based on the extracted parameters, the corrected beam is controlled to perform a multi-line global scanning operation along a preset crystal orientation. Through multi-line global scanning, a continuously bridged micro-explosion modification layer is formed inside the crystal of a 4H-SiC wafer.

[0012] In one implementation of the first aspect of the present invention, the process of collecting wafer quality indicators after stripping and expanding the process database or triggering incremental learning of a neural network based on the wafer quality indicators includes: The surface roughness and critical peel fracture strength of the wafer after peeling were measured using testing equipment. The surface roughness and critical peel fracture strength obtained by the measurement are compared with the preset qualified threshold set by the system. When the surface roughness and the critical peel fracture strength simultaneously meet the preset qualified threshold, the wafer physical properties corresponding to the batch of wafers and the currently used laser slicing process parameter set are jointly packaged into a complete sample, and the complete sample is added and stored in the process database. When either the surface roughness or the critical peel fracture strength fails to meet the preset qualified threshold, the data corresponding to all current working conditions will be marked and included in the calibration dataset. The calibration dataset is invoked to trigger the neural network to perform an online incremental learning process, which updates the network weights within the neural network and reduces the prediction bias of subsequent process parameters.

[0013] In one implementation of the first aspect of the present invention, before obtaining the resistivity, doping type, and target focal depth of the 4H-SiC wafer to be sliced, the process database is further constructed: We selected 4H-SiC wafer samples with different doping types and collected the resistivity and doping concentration of each wafer sample. Multiple sets of different target focal depth conditions were set up, and multiple sets of laser single-line micro-explosion and full-domain scanning experiments were carried out for each wafer sample. Simulation calculations were performed using a dual-temperature model to obtain information on lattice temperature evolution, and the actual peeling quality test results of the wafers were obtained after the experiment. Based on the criteria for acceptance of lattice temperature, wafer peel fracture strength, and wafer surface roughness, multiple qualified processing samples were selected from all test results. Each qualified processing sample is organized into a standard sample pair. The standard sample pair uses doping type, resistivity, and target focal depth as input dimensions, and laser power, scanning speed, scanning step size, and pulse repetition frequency as output dimensions. Based on the statistical calculation of all the standard samples, the mean log resistivity and the standard deviation of log resistivity are obtained. The statistically obtained constants are then stored in the process database for retrieval when generating the continuous backbone input features, thus completing the overall construction of the process database.

[0014] Secondly, the present invention provides a 4H-SiC wafer intelligent laser slicing system.

[0015] A 4H-SiC wafer intelligent laser slicing system includes a parameter acquisition and processing module, an intelligent process prediction and control system, an ultrafast laser modulation and focusing module, a motion platform and Z-axis closed-loop tracking module, a stress-guided peeling unit, and a slicing quality detection and closed-loop feedback unit. The parameter acquisition and processing module is used to acquire the resistivity, doping type, and target focal depth of the 4H-SiC wafer to be processed, perform logarithmic transformation and standardization on the resistivity to generate continuous backbone input features, perform one-hot encoding and splicing on the doping type and the target focal depth to generate discrete condition vectors, and transmit the continuous backbone input features and the discrete condition vectors to the intelligent process prediction and control system. The intelligent process prediction and control system has a built-in linear modulation neural network for receiving the continuous backbone input features and the discrete condition vector, inferring and outputting a laser slicing process parameter set, and outputting the laser slicing process parameter set to the ultrafast laser control and focusing module and the motion platform and Z-axis closed-loop tracking module. It also receives the wafer quality indicators after stripping from the slicing quality detection and closed-loop feedback unit, and performs process database expansion or neural network incremental learning.

[0016] In one implementation of the second aspect of the present invention, the ultrafast laser modulation and focusing module includes an ultrafast laser, a half-wave plate, a polarizer, a beam expander, a liquid crystal spatial light modulator, a first lens, a second lens, a dichroic beam splitter, a coaxial CCD camera, and a micro-focusing objective lens equipped with a piezoelectric ceramic axial fine adjustment mechanism. The processing laser output from the ultrafast laser passes sequentially through the half-wave plate and the polarizer to adjust the laser polarization state and output power. The adjusted beam is then incident on the beam expander. The beam expander collimates and magnifies the incident beam, and the processed beam is then incident on the liquid crystal spatial light modulator. The liquid crystal spatial light modulator loads the spherical aberration compensation phase map output from the intelligent process prediction and control system to perform phase correction on the beam to obtain a corrected beam. After phase correction, the corrected beam is output and passes sequentially through the 4F optical relay imaging system composed of the first lens and the second lens. The beam processed by the 4F optical relay imaging system is incident on the dichroic beam splitter. The dichroic beam splitter reflects the corrected beam. The reflected corrected beam is incident on the micro-focusing objective. The micro-focusing objective focuses the corrected beam and projects it into the crystal interior of the 4H-SiC wafer to be processed. The dichroic beam splitter simultaneously transmits the observation light path from the wafer end, and the transmitted observation light path is transmitted to the coaxial CCD camera. The coaxial CCD camera relies on the transmitted light path to complete the coaxial alignment observation of the microscopic features of the wafer surface and the focal length of the spot. The piezoelectric ceramic axial fine adjustment mechanism receives the height control signal output from the motion platform and the Z-axis closed-loop tracking module, and drives the microscope focusing objective to complete the dynamic adjustment of its axial position.

[0017] In one implementation of the second aspect of the present invention, the motion platform and the Z-axis closed-loop tracking module include a vacuum chuck, a horizontal rotation axis, a three-axis linear feed mechanism, and a coaxial laser displacement sensor. The vacuum chuck is used to adsorb and fix the 4H-SiC wafer to be processed, ensuring that the wafer position does not shift during the scanning process. The horizontal rotation axis is used to drive the wafer to rotate and complete the alignment operation between the specific crystal orientation of the wafer and the scanning feed direction. The three-axis linear feed mechanism is used to carry the wafer and complete the high-speed scanning feed and the feed movement in the step direction during the processing. The coaxial laser displacement sensor collects the surface height fluctuation data of the wafer in real time and converts the collected surface height fluctuation data of the wafer into a height signal and outputs it to the intelligent process prediction and control system. The stress-guided peeling unit includes upper and lower opposing tension clamps. The upper and lower opposing tension clamps can apply uniform tensile stress to the 4H-SiC wafer to be processed, which has a micro-explosion modified layer already formed inside, and induce transverse micro-cracks inside the wafer to expand and penetrate along the modified layer, so as to obtain the peeled wafer sheet and realize the peeling of the wafer sheet. The slice quality detection and closed-loop feedback unit is used to measure the surface roughness and critical peeling tensile force of the peeled wafer after peeling is completed, and to obtain the wafer quality index after peeling. When the detection indicators meet the qualification conditions, the slice quality detection and closed-loop feedback unit will send back the relevant data of the qualified sample corresponding to the quality indicators of the peeled wafer to the intelligent process prediction and control system to complete the database expansion. When the detection index exceeds the acceptable range, the slice quality detection and closed-loop feedback unit outputs a recalibration signal to the intelligent process prediction and control system, triggering the online incremental update and adaptive evolution of the neural network.

[0018] Compared with the prior art, the beneficial effects of the present invention are: This invention acquires the resistivity, doping type, and target focal depth of a 4H-SiC wafer to be sliced, and generates continuous backbone input features and discrete conditional vectors, respectively. Using a conditional neural network, the features are subjected to channel-by-channel affine modulation to output a set of laser slicing process parameters adapted to the current wafer conditions. This allows for matching processing parameters to the material properties of wafers with different doping attributes, alleviating the difficulty in matching processing conditions for wafers with different doping types. Simultaneously, a spherical aberration compensation phase diagram is calculated based on the target focal depth to correct beam focusing distortion, resulting in a corrected beam. The corrected beam is then used for internal wafer modification scanning, reducing the interference of beam focusing deviation caused by the crystal medium on the internal modification effect. After wafer stress stripping, post-stripping wafer quality indicators are collected. These indicators are used to expand the process database or trigger incremental learning of the neural network, allowing the actual slicing quality feedback to influence subsequent parameter inference processes, gradually improving the accuracy of the neural network's inference of process parameters under different conditions. Therefore, this invention can adapt to laser slicing processing scenarios for 4H-SiC wafers with different doping types, improving processing adaptation and helping to reduce defects during wafer processing.

[0019] This invention constructs a modulation paradigm of "continuous physical properties - discrete doping conditions," solving the problem that traditional deep learning struggles to fit strongly nonlinear abrupt changes. This invention explicitly designs a conditional neural network based on the Feature Linear Modulation (FiLM) mechanism, using log-normalized resistivity as the main continuous input and doping type and depth of focus as conditional vectors. This is achieved by implementing channel-by-channel scaling in the hidden layers. ) and offset ( Affine transformation enables adaptive and precise decoupling of nonlinear processing boundaries under different doping types, significantly improving the algorithm's ability to characterize and express the physical properties of complex semiconductor material processing.

[0020] This invention overcomes the limitations of empirical trial and error, achieving "second-level reverse prediction and zero-trial-error decision-making" for cross-batch / cross-doping type slicing process parameters. Traditional laser slicing process adjustments rely entirely on empirical rules or time-consuming, laborious gridded parameter experiments, which can easily lead to the scrapping of large quantities of expensive substrates when faced with wafers exhibiting resistivity fluctuations. This invention utilizes a data-driven reverse mapping network to establish a model from "initial wafer electrical / structural properties" to "laser power P, scanning speed v, and scanning step size." repetition frequency The prediction mechanism, and the full parameter decision coefficients of the test set. This enables second-level adaptive intelligent generation of process parameters for new batches or wafers with different doping levels, completely eliminating the need for trial and error based on human experience in traditional processing and greatly shortening the process development cycle for new batches of wafers.

[0021] This invention establishes a "single unified model" to replace the "fragmented patchwork of multiple models," possessing extremely strong model generalization and cross-material transfer capabilities. Existing technologies, to handle multiple doping types, typically require training and maintaining separate prediction models for each doping type, resulting in high maintenance costs. This invention introduces discrete doping characteristics as conditional branches into a neural network, utilizing a unified network paradigm to integrate the processing rules of all doping types, including semi-insulating (high transmittance / low carrier) and conductive (high absorption / strong path dissipation). This model structure not only adapts to minute resistivity fluctuations within the same doping type but also seamlessly transfers to laser micro / nano fabrication of other wide-bandgap semiconductors such as GaN and diamond, where doping absorption abruptly occurs.

[0022] This invention constructs an intelligent laser slicing system integrating "attribute acquisition - deep algorithm inference - laser system execution". This invention deeply integrates a deep learning process prediction engine with detection data (resistivity, focusing depth) and laser processing parameter control. The process parameters output by the neural network model through inverse reasoning can be directly converted into industrial control commands, realizing "material property input". Algorithmic reasoning and decision making The fully automated intelligent processing closed loop of "laser adaptive compensation scanning" provides core deep learning algorithm support for the transformation of silicon carbide substrate processing equipment from "experience-controlled" to "data-driven intelligent control".

[0023] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0024] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0025] Figure 1 A schematic flowchart of a smart laser slicing method for 4H-SiC wafers provided as an exemplary embodiment of the present invention; Figure 2 A schematic diagram of a FiLM neural network provided as an exemplary embodiment of the present invention; Figure 3 A schematic diagram of the optical path of a 4H-SiC wafer intelligent laser slicing system provided as an exemplary embodiment of the present invention; The components include: 1. Ultrafast laser; 2. Half-wave plate; 3. Polarizer; 4. Beam expander; 5. Liquid crystal spatial light modulator; 6. First lens; 7. Second lens; 8. Coaxial CCD camera; 9. Dichroic beam splitter; 10. Microscopic focusing objective equipped with a piezoelectric ceramic axial fine-tuning mechanism; 11. Motion platform and Z-axis closed-loop tracking module; 12. Intelligent process prediction and control system; 13. Parameter acquisition and processing module; 14. 4H-SiC wafer to be processed; 15. Wafer sheet that has been peeled off; 16. Stress-guided peeling unit; 17. Slicing quality inspection and closed-loop feedback unit. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0027] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0028] Silicon carbide (4H-SiC), as a representative substrate material for third-generation wide-bandgap semiconductors, has irreplaceable strategic value in fields such as new energy vehicles, high-voltage power grids, and radio frequency devices. However, 4H-SiC has a high Mohs hardness of 9.5 and low fracture toughness, resulting in large kerf losses when cut with traditional diamond wire saws. This technology addresses severe defects such as deep subsurface damage (SSD). Ultrafast laser internal modification slicing utilizes nonlinear absorption effects to focus near-infrared pulsed lasers inside the crystal, inducing multiphoton ionization and forming a high-density micro-explosion modified layer. Subsequent stress-guided peeling achieves sheet separation, offering significant advantages such as non-contact processing, zero kerf loss, and low damage.

[0029] However, it should be noted that in industrial semiconductor manufacturing, 4H-SiC substrates are classified into several types according to electrical requirements, including high-purity semi-insulating type, vanadium-doped semi-insulating type, nitrogen-doped N-type, and aluminum-doped P-type. Different doping types can cause the resistivity of the material to span nearly 13 orders of magnitude. This leads to the material's sensitivity to near-infrared bands (such as...). The light absorption mechanism of laser light undergoes a drastic change with the initial free carrier concentration. Among these, semi-insulating wafers with extremely low free carrier concentration exhibit extremely high transmittance (absorption rate). This highly localizes light energy to the focal plane, making it prone to excessive micro-explosions and deep longitudinal cracks under conventional fixed power, resulting in a sharp increase in the roughness of the peeled surface; while conductive wafers containing a high concentration of initial free carriers exhibit strong path light absorption of near-infrared light (absorption rate reaching...). This results in severe dissipation of the laser light as it penetrates the incident path, making it difficult for the effective energy reaching the focal plane to cross the micro-explosion modification threshold. This leads to insufficient bridging of the modified layer, a sharp increase in peeling force, and edge breakage. Traditional industrial processing relies on fixed parameters or manual trial and error, which cannot achieve adaptive control when dealing with wafers with varying resistivity across different doping batches or even within the same batch, and easily leads to the scrapping of large numbers of wafers. This invention establishes an adaptive mapping and control mechanism of "resistivity / doping type - path light absorption attenuation - focal plane modification threshold," thereby achieving intelligent slicing across doping types in complex laser peeling tasks involving various types of silicon carbide. This invention adopts a "fusion of physical mechanism-driven and data-driven" approach, designing a conditional neural network based on Feature Linear Modulation (FiLM). The algorithm uses the continuous physical properties of the wafer (logarithm of resistivity) as the basis for the algorithm. Using the main input as the parameter set and discrete structural / attribute constraints (doping type, target focal depth, etc.) as the conditional modulation vector, the system reversely and accurately outputs the parameter set required for the laser slicing process. .

[0030] Before training the FiLM conditional neural network, a training database covering laser slicing processes with multiple doping types was pre-constructed. First, for various 4H-SiC wafer samples, including high-purity semi-insulating, vanadium-doped semi-insulating, nitrogen-doped N-type, and light aluminum-doped P-type, the room-temperature resistivity of each wafer was measured using a non-contact resistivity testing system, secondary ion mass spectrometry (SIMS), and spectrophotometer. Doping concentration and near-infrared light absorption rate; Subsequently, at different target depths of focus, the laser power was considered. Scanning speed Feed step Pulse repetition frequency Laser single-line micro-explosion and multi-line full-domain scanning experiments were carried out sequentially. Combined with a two-temperature model (TTM), the carrier excitation and lattice temperature evolution at the focal plane were simulated and calculated. Based on the simulation results and the tensile peel fracture strength and surface roughness of the actual peeled wafer, the lattice temperature was used to reach the thermal decomposition threshold. ), peel strength And the surface roughness of the peeling surface As a qualification criterion, multiple sets of qualified processing parameter samples were selected; The qualified samples were then organized into categories based on [doping type, resistivity] The target focal depth is used as the input feature vector, and the laser power is used as the input feature vector. repetition frequency Scanning speed Scan step distance 】 represents the standard sample data pairs for the target output vector.

[0031] Then, the common logarithm of the resistivity of all samples in the database is taken. Statistical calculations yielded the global logarithmic resistivity mean and standard deviation: (1); And and As static calibration constants, they are permanently stored in the host computer system and parameter preprocessing module for online inference and the generation of standardized tensors when the model is divided into training and test sets in an 8:2 ratio.

[0032] Specifically, such as Figure 1 As shown, this invention provides a smart laser slicing method for 4H-SiC wafers, comprising the following steps: Step S101: Acquisition of wafer electrical and structural features and data preprocessing.

[0033] The room temperature resistivity of 4H-SiC material varies depending on the doping elements and their concentrations. Spanning nearly 13 orders of magnitude (conductivity type approximately) Semi-insulating type If you directly input the raw value The value is The semi-insulating wafer will produce a dominant, huge gradient, with a value of The conductive wafer characteristics will be completely ignored in floating-point operations (numerical underflow), leading to backpropagation gradient explosion or model failure to converge.

[0034] Perform common logarithmic transformations This transforms the nonlinear multiplicative exponential relationship into an additive linear relationship. Smooth mapping of physical intervals to The linearly comparable interval. Then, the pre-calibrated logarithmic mean from the benchmark process database is called. With log standard deviation Perform a Z-score spatially standardized linear transformation: (2); Generate dimensionless 1D continuous backbone input features with zero mean and unit variance. It is used as input to the hidden layer of the backbone of the FiLM conditional neural network to participate in the extraction of basic continuous electrical mapping features.

[0035] Obtain discrete structure / property constraints, including wafer doping type (high-purity semi-insulator, vanadium-doped semi-insulator, nitrogen-doped N-type, aluminum-doped P-type) and preset target depth of focus (e.g. (Interval gradient).

[0036] Subsequently, the parameter acquisition and processing module performs one-hot encoding on the discrete features, mapping the four doping types into four-dimensional orthogonal one-hot vectors. ,Will Each target depth-of-focus level is mapped as follows: 3D orthogonal one-hot vector: (3); Furthermore, the aforementioned and Perform vector concatenation to generate a vector with dimension 1. Discrete conditional vector: (4); Step S102: Train the FILM neural network and intelligently predict process parameters.

[0037] Preprocessed continuous features The input is fed into the backbone feature extraction network of the FiLM conditional neural network (containing at least two fully connected hidden layers, each layer having a set feature extraction layer). (using a ReLU activation function) 10 neurons to learn the basic continuous mapping between resistivity and modification parameters.

[0038] condition vector The input is fed into the conditional modulation branch network, and a scaling factor vector with the same dimension as the hidden features of the backbone is generated through nonlinear mapping. With offset factor vector : (5); (6); For the backbone network Hidden feature representation of layers Perform channel-by-channel affine modulation transform: (7); In the formula, The model is an Adama product. It is trained and optimized using Smooth L1 Loss (Huber Loss), and finally outputs the inversely normalized optimal laser slicing process parameter set, including laser power. Scanning speed Scan step distance and pulse repetition frequency .

[0039] like Figure 2 The diagram shown is a schematic of a specific FiLM neural network, which is divided into three parts: feature input, backbone network and FiLM modulation, and output parameters. In the feature input part, resistivity... The continuous backbone input features were obtained through logarithmic and standardized operations. Different wafer types Depth of Target Focus After processing, discrete conditional vectors are obtained. The two types of features are then fed into the backbone network and the conditional modulation branch, respectively. The backbone network performs feature extraction operations on the continuous backbone input features to obtain the backbone hidden features. The conditional modulation branch performs a linear mapping operation on the discrete conditional vectors and outputs scaling factors respectively. With offset factor According to formula (7), the channel-by-channel affine modulation operation is completed to obtain the modulated hidden features; the modulated hidden features are sent to the output parameter module for inference and output respectively. Corresponding laser power , Corresponding repetition frequency , Corresponding scanning speed , Corresponding scan interval This enables the prediction and output of laser slicing process parameters based on wafer electrical properties, doping type, and target focusing depth.

[0040] Step S103: Calculate the actual focusing depth and dynamic correction of optical spherical aberration.

[0041] According to air ( ) and 4H-SiC crystal ( Based on the refractive index difference and target depth of focus, the actual focal shift is calculated using the refractive mapping formula: (8); The calculated spherical aberration-corrected phase map is loaded into the spatial light modulator to correct the focal spot distortion when the beam penetrates the medium. Simultaneously, a coaxial height sensor is used to monitor the wafer surface height fluctuations in real time, and this is used in conjunction with a piezoelectric ceramic driving objective to achieve high-precision Z-axis closed-loop height compensation.

[0042] Step S104: Perform internal modification scanning on the differentially doped wafer through a laser parameter compensation and control mechanism.

[0043] Make the laser beam along the carbon surface ( (Surface) perpendicularly incident and along a predetermined crystal orientation (such as) Multi-line global scanning is performed. During the scanning process, a differentiated energy control strategy predicted in step two is executed to precisely generate a continuously bridged micro-explosion modification layer inside the wafer focal plane.

[0044] When performing step four above, the control parameters predicted inversely by the neural network strictly follow the laser parameter compensation mechanism for different doping types: for semi-insulating wafers (high transmittance / low carriers), due to the material absorption rate... With very few free carriers, a low laser power is predicted. (like ) with higher scanning speed (like This combination can suppress excessive micro-explosions and ionizing avalanches, effectively limiting longitudinal damage. Energy compensation strategy for conductive wafers (high absorption / high carrier): Due to the high light absorption rate along the material path... Predicting higher laser power (like This strategy is used to dynamically compensate for the light energy dissipated by free carriers in the transmission path, ensuring that the energy density focused to the focal plane is still sufficient to push the lattice temperature to the thermal decomposition threshold. This ensures that the modified layer is fully bridged.

[0045] Step S105: Stress-guided low-damage wafer separation.

[0046] Applying uniform mechanical tensile stress or thermal stress to the 4H-SiC wafer that has undergone internal scanning modification induces internal transverse microcracks to propagate at a fixed height along the modified layer, thereby achieving high-quality peeling and separation of the wafer sheet.

[0047] Step S106: Iterative process database and closed-loop self-evolution of neural network based on wafer quality feedback after stripping.

[0048] After completing the 4H-SiC wafer peeling and separation, a closed-loop self-evolutionary mechanism based on post-evaluation of slice quality and dynamic expansion of the process database was constructed: the surface roughness of the peeled surface of the separated sheet was measured by a three-dimensional optical profilometer, a laser confocal microscope, and a tensile force sensor. ), longitudinal damage depth and critical peel fracture strength, and compared with a preset quality qualification threshold (preferred). And peel strength Comparative analysis was performed. When the slice quality meets the standard, the system automatically applies the currently executed optimal laser slicing process parameters (laser power). Scanning speed Scan step distance Pulse frequency ) and the initial physical properties (resistivity) of this batch of wafers The training set is packaged into standard sample data pairs (doping type, target focal depth, etc.), and then sent back and appended to the process database to achieve autonomous dynamic expansion of the training set. When the slice quality fails to meet the standard or the measured indicators fluctuate and drift, the system automatically marks the current working condition data and stores it in the calibration dataset, triggering the online incremental learning and weight fine-tuning of the FiLM conditional neural network. By re-optimizing the mapping parameters of the conditional modulation branch network, the prediction deviation caused by material batch fluctuations or unknown impurities is eliminated, realizing the continuous dynamic accumulation of the slice process database and the closed-loop adaptive iterative evolution of the neural network prediction model.

[0049] This invention also proposes a laser intelligent slicing system for implementing the above method, comprising a parameter acquisition and processing module 13, an intelligent process prediction and control system 12, an ultrafast laser modulation and focusing module, a motion platform and Z-axis closed-loop tracking module 11, a stress-guided ablation unit 16, and a slice quality detection and closed-loop feedback unit 17. Parameter acquisition and processing module 13: Used to obtain the matrix resistivity of the 4H-SiC wafer to be sliced. The doping type and target focal depth are determined, and the resistivity is logarithmically transformed and normalized to generate continuous input features. One-hot encoding is performed on the doping type and target focal depth to generate discrete condition vectors. ; The intelligent process prediction and control system 12 based on the FiLM neural network is connected to the parameter acquisition and processing module 13 and has a built-in trained Feature Linear Modulation (FiLM) conditional neural network for predicting and controlling processes based on input parameters. and Inverse adaptive reasoning outputs optimal process parameter instructions This includes laser power P, pulse repetition frequency f, scanning speed v, and scanning step size. It also features process database management, data visualization, and online incremental learning of models based on post-hoc feedback of slice quality. Ultrafast laser modulation and focusing module: Controlled by the host computer control system, it includes an ultrafast laser, a polarization power adjustment component, a beam expander system, a liquid crystal spatial light modulator (LC-SLM), a 4F optical relay system, and a focusing objective lens equipped with a piezoelectric ceramic (PZT) fine-tuning mechanism. After the spatial light modulator dynamically loads the medium spherical aberration compensation phase map, it outputs a high-quality focused beam with path light energy attenuation compensation to the wafer focal plane. Motion platform and Z-axis closed-loop tracking module 11: includes a vacuum chuck for supporting the wafer and a horizontal angle rotation for calibrating the wafer's crystal orientation. Shaft mechanism, high precision A three-axis linear motion mechanism and a coaxial laser displacement sensor; via rotation The axis enables the wafer to The crystal orientation is precisely aligned with the X-axis scanning direction, and during the high-speed X-axis scanning and Y-axis step feed, a coaxial laser displacement sensor is used to collect real-time data on the height fluctuations of the wafer surface. The Z-axis height of the piezoelectric ceramic fine-tuning objective is driven in a closed loop to ensure that the internal focus is absolutely fixed at a certain depth. Stress-guided peeling unit 16: includes upper and lower opposing clamps and a force measuring mechanism, used to apply uniform tensile stress to the wafer with a continuous modified layer inside, and induce transverse microcracks to fully propagate along the modified layer at a fixed height to achieve low-damage separation of the sheet. Slice quality inspection and closed-loop feedback unit 17: used to determine the surface roughness Ra and critical peeling tensile force F of the sheet after peeling, and to transmit the qualified working condition data as new samples back to expand the upper computer process database, or to trigger the online incremental training and weight adaptive iterative update of the FiLM neural network in the upper computer when the quality exceeds the standard.

[0050] like Figure 2 The diagram shows a specific optical path. An ultrafast laser 1 emits near-infrared ultrafast pulsed laser light. A half-wave plate 2 and a polarizer 3 together form a precision adjustment assembly for laser polarization and output power. A beam expander 4 is used for beam collimation and aperture magnification to match the optical modulation device. A liquid crystal spatial light modulator 5 dynamically loads a phase map to compensate for spherical aberration, eliminating focusing aberrations caused by the beam penetrating a high-refractive-index crystal medium. A first lens 6 and a second lens 7 together form a 4F optical repeater. The system includes: a coaxial CCD camera 8 for coaxial alignment observation of microscopic features on the wafer surface and the focal length of the light spot; a dichroic beam splitter 9 for reflecting the processing laser and transmitting the observation light path; a micro-focusing objective 10, equipped with a piezoelectric ceramic axial fine-tuning mechanism, to precisely focus the shaped laser to a preset depth inside the wafer, forming a micro-explosion modification area; and a high-precision motion platform and Z-axis closed-loop tracking module 11, including a vacuum chuck, a horizontal rotation axis, and a three-axis linear feed mechanism, to complete crystal orientation alignment, high-speed scanning, and dynamic depth determination of the wafer surface height.

[0051] The intelligent process prediction and control system 12 based on the FiLM neural network serves as the control center of the entire system. It outputs parameters such as laser power, scanning speed, step size, and repetition frequency based on the input physical characteristics through reverse adaptive reasoning. This provides control commands to the ultrafast laser 1, the high-precision motion platform, and the Z-axis closed-loop tracking module 11. The parameter acquisition and processing module 13 acquires the continuous resistivity, discrete doping type, and target depth of focus of the wafer to be processed. After logarithmic normalization and one-hot encoding conversion, the data is transmitted to the intelligent process prediction and control system 12 based on the FiLM neural network. The 4H-SiC wafer 14 to be processed, after its physical characteristics are acquired, is placed on the high-precision motion platform. Above the platform and the Z-axis closed-loop tracking module 11, there is the original wafer substrate to be sliced; the peeled wafer 15 is the wafer obtained after slicing; the stress-guided peeling unit 16 includes upper and lower opposing tension jigs, which apply uniform tensile stress to the modified wafer to induce complete penetration of transverse microcracks and realize wafer peeling; the slicing quality detection and closed-loop feedback unit 17 is used to measure the surface roughness and critical peeling tensile force of the peeled wafer 15 and store the compliant working condition data into the database; when the detection indicators are abnormal, a recalibration signal is sent to the intelligent process prediction and control system 12 based on the FiLM neural network to trigger the online incremental update and adaptive evolution of the neural network.

[0052] The following components work together to construct a complete intelligent slicing processing system: 1. Ultrafast laser; 2. Half-wave plate; 3. Polarizer; 4. Beam expander; 5. Liquid crystal spatial light modulator; 6. First lens; 7. Second lens; 8. Coaxial CCD camera; 9. Dichroic beam splitter; 10. Microscopic focusing objective; 11. High-precision motion platform and Z-axis closed-loop tracking module; 12. Intelligent process prediction and control system based on FiLM neural network; 13. Parameter acquisition and processing module; 14. 4H-SiC wafer to be processed; 15. Wafer stripped off; 16. Stress-guided stripping unit; 17. Slicing quality detection and closed-loop feedback unit.

[0053] More specifically, the ultrafast laser modulation and focusing module includes an ultrafast laser 1, a half-wave plate 2, a polarizer 3, a beam expander 4, a liquid crystal spatial light modulator 5, a first lens 6, a second lens 7, a dichroic beam splitter 9, a coaxial CCD camera 8, and a microscopic focusing objective lens 10 equipped with a piezoelectric ceramic axial fine adjustment mechanism. The processing laser output from the ultrafast laser 1 passes sequentially through the half-wave plate 2 and the polarizer 3 to adjust the laser polarization state and output power. After adjustment, the beam is incident on the beam expander 4. The beam expander 4 performs beam collimation and aperture magnification on the incident beam. After processing, the beam is incident on the liquid crystal spatial light modulator 5. The liquid crystal spatial light modulator 5 loads the spherical aberration compensation phase map output from the intelligent process prediction and control system 12 to perform phase correction on the beam to obtain a corrected beam. After phase correction, the corrected beam is output and passes sequentially through the 4F optical relay imaging system composed of the first lens 6 and the second lens 7. The beam processed by the 4F optical relay imaging system is incident on the dichroic beam splitter 9. The dichroic beam splitter 9 reflects the corrected beam. The reflected corrected beam is incident on the micro-focusing objective 10. The micro-focusing objective 10 focuses the corrected beam and projects it into the crystal interior of the 4H-SiC wafer 14 to be processed. The dichroic beam splitter 9 simultaneously transmits the observation light path from the wafer end, and the transmitted observation light path is transmitted to the coaxial CCD camera 8. The coaxial CCD camera 8 relies on the transmitted light path to complete the coaxial alignment observation of the micro-features of the wafer surface and the focal length of the light spot. The piezoelectric ceramic axial fine-tuning mechanism receives the height control signal output from the high-precision motion platform and the Z-axis closed-loop tracking module 11, thereby driving the microscope focusing objective 10 to complete the dynamic adjustment of its axial position.

[0054] The high-precision motion platform and Z-axis closed-loop tracking module 11 include a vacuum chuck, a horizontal rotation axis, a three-axis linear feed mechanism, and a coaxial laser displacement sensor. The vacuum chuck is used to adsorb and fix the 4H-SiC wafer 14 to be processed, ensuring that the wafer position does not shift during the scanning process. The horizontal rotation axis is used to drive the wafer to rotate and complete the alignment operation between the specific crystal orientation of the wafer and the scanning feed direction. The three-axis linear feed mechanism is used to carry the wafer and complete the high-speed scanning feed and the feed movement in the step direction during the processing. The coaxial laser displacement sensor collects the surface height fluctuation data of the wafer in real time and converts the collected surface height fluctuation data of the wafer into a height signal and outputs it to the intelligent process prediction and control system 12. The stress-guided peeling unit 16 includes upper and lower opposing tension clamps. The upper and lower opposing tension clamps can apply uniform tensile stress to the 4H-SiC wafer 14 to be processed, which has a micro-explosion modified layer already formed inside, and induce transverse micro-cracks inside the wafer to expand and penetrate along the modified layer, so as to obtain the peeled wafer sheet 15 and realize the peeling of the wafer sheet. The slice quality detection and closed-loop feedback unit 17 is used to measure the surface roughness and critical peeling pull of the peeled wafer 15 after peeling is completed, and obtain the wafer quality index after peeling; when the detection index meets the qualified conditions, the slice quality detection and closed-loop feedback unit 17 transmits the relevant data of the qualified sample corresponding to the wafer quality index after peeling back to the intelligent process prediction and control system 12 to complete the database expansion. When the detection index exceeds the acceptable range, the slice quality detection and closed-loop feedback unit 17 outputs a recalibration signal to the intelligent process prediction and control system 12, triggering the online incremental update and adaptive evolution of the neural network.

[0055] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent laser slicing of 4H-SiC wafers, characterized in that, Includes the following processes: Obtain the resistivity, doping type, and target focal depth of the 4H-SiC wafer to be sliced; Logarithmic transformation and standardization are performed on the resistivity to generate continuous backbone input features. The doping type and the target focal depth are encoded using one-hot encoding and then concatenated to generate a discrete conditional vector. The continuous backbone input features are input into the backbone network to obtain the backbone hidden features. The discrete conditional vector is used to generate scaling factors and offset factors. The scaling factors and offset factors are used to perform channel-wise affine modulation on the backbone hidden features to obtain modulated hidden features. The set of laser slicing process parameters is output based on the modulated hidden features. Based on the target depth of focus, the focal shift is calculated to generate a spherical aberration compensation phase diagram. Based on the spherical aberration compensation phase diagram, the beam focusing distortion is corrected to obtain the corrected beam. Based on the laser slicing process parameter set, the corrected beam is used to perform internal wafer modification scanning, and stress is applied to the modified wafer to achieve wafer peeling. Collect wafer quality indicators after stripping, and expand the process database or trigger incremental learning of neural networks based on the wafer quality indicators after stripping.

2. The intelligent laser slicing method for 4H-SiC wafers as described in claim 1, characterized in that, The resistivity is logarithmically transformed and standardized to generate continuous backbone input features, including: The logarithmic resistivity is obtained by taking the common logarithm of the resistivity. Retrieve the mean log resistivity and standard deviation log resistivity pre-obtained from a large number of wafer samples within the process database; Subtract the mean of the logarithmic resistivity from the logarithmic resistivity, and then divide by the standard deviation of the logarithmic resistivity to complete the standardized calculation of zero mean unit variance. The dimensionless continuous backbone input features are obtained through the aforementioned standardized calculations. These continuous backbone input features are used to input the backbone network for subsequent feature extraction operations.

3. The intelligent laser slicing method for 4H-SiC wafers as described in claim 1, characterized in that, Generating scaling and offset factors using the discrete condition vector includes: The discrete conditional vector is completely input into the conditional modulation branch network; Two sets of independent linear mapping operations are performed within the conditional modulation branch network. The scaling factor is output through one of the sets of linear mapping operations; The offset factor is output through another set of linear mapping operations; The vector dimensions of the scaling factor and the offset factor obtained from the output are consistent with the dimensions of the hidden features of the backbone network output.

4. The intelligent laser slicing method for 4H-SiC wafers as described in claim 1, characterized in that, The trunk hidden features are obtained by performing channel-wise affine modulation using the scaling factor and the offset factor, including: The scaling factor is multiplied element-wise with the trunk hidden feature according to the channel dimension; Obtain the intermediate result of the element-wise multiplication operation; The intermediate calculation result is then added element by element to the offset factor. After performing element-by-element addition, the modulated hidden feature is output. The modulated hidden features incorporate doping and depth-of-focus constraint information from the discrete condition vector, which are used for subsequent inference calculations to obtain the laser slicing process parameter set.

5. The intelligent laser slicing method for 4H-SiC wafers as described in claim 1, characterized in that, Based on the laser slicing process parameter set, the corrected beam is used to perform an internal wafer modification scan, including: Real-time acquisition of wafer surface height undulation data; The wafer surface height fluctuation data is converted into a control signal, and the axial height of the objective lens is adjusted in a closed loop to maintain the stability of the actual depth of the laser focus inside the wafer. The laser slicing process parameter set was analyzed to extract the laser power, scanning speed, scanning step size, and pulse repetition frequency. Based on the extracted parameters, the corrected beam is controlled to perform a multi-line global scanning operation along a preset crystal orientation. Through multi-line global scanning, a continuously bridged micro-explosion modification layer is formed inside the crystal of a 4H-SiC wafer.

6. The intelligent laser slicing method for 4H-SiC wafers as described in claim 1, characterized in that, Collect wafer quality indicators after stripping, and expand the process database or trigger incremental learning of the neural network based on the wafer quality indicators after stripping, including: The surface roughness and critical peel fracture strength of the wafer after peeling were measured using testing equipment. The surface roughness and critical peel fracture strength obtained by the measurement are compared with the preset qualified threshold set by the system. When the surface roughness and the critical peel fracture strength simultaneously meet the preset qualified threshold, the wafer physical properties corresponding to the batch of wafers and the currently used laser slicing process parameter set are jointly packaged into a complete sample, and the complete sample is added and stored in the process database. When either the surface roughness or the critical peel fracture strength fails to meet the preset qualified threshold, the data corresponding to all current working conditions will be marked and included in the calibration dataset. The calibration dataset is invoked to trigger the neural network to perform an online incremental learning process, which updates the network weights within the neural network and reduces the prediction bias of subsequent process parameters.

7. The intelligent laser slicing method for 4H-SiC wafers as described in claim 1, characterized in that, Before obtaining the resistivity, doping type, and target focal depth of the 4H-SiC wafer to be sliced, the process database is constructed. We selected 4H-SiC wafer samples with different doping types and collected the resistivity and doping concentration of each wafer sample. Multiple sets of different target focal depth conditions were set up, and multiple sets of laser single-line micro-explosion and full-domain scanning experiments were carried out for each wafer sample. Simulation calculations were performed using a dual-temperature model to obtain information on lattice temperature evolution, and the actual peeling quality test results of the wafers were obtained after the experiment. Based on the criteria for acceptance of lattice temperature, wafer peel fracture strength, and wafer surface roughness, multiple qualified processing samples were selected from all test results. Each qualified processing sample is organized into a standard sample pair. The standard sample pair uses doping type, resistivity, and target focal depth as input dimensions, and laser power, scanning speed, scanning step size, and pulse repetition frequency as output dimensions. Based on the statistical calculation of all the standard samples, the mean log resistivity and the standard deviation of log resistivity are obtained. The statistically obtained constants are then stored in the process database for retrieval when generating the continuous backbone input features, thus completing the overall construction of the process database.

8. A 4H-SiC wafer intelligent laser slicing system, characterized in that, It includes a parameter acquisition and processing module, an intelligent process prediction and control system, an ultrafast laser control and focusing module, a motion platform and Z-axis closed-loop tracking module, a stress-guided peeling unit, and a slice quality detection and closed-loop feedback unit. The parameter acquisition and processing module is used to acquire the resistivity, doping type, and target focal depth of the 4H-SiC wafer to be processed, perform logarithmic transformation and standardization on the resistivity to generate continuous backbone input features, perform one-hot encoding and splicing on the doping type and the target focal depth to generate discrete condition vectors, and transmit the continuous backbone input features and the discrete condition vectors to the intelligent process prediction and control system. The intelligent process prediction and control system has a built-in linear modulation neural network for receiving the continuous backbone input features and the discrete condition vector, inferring and outputting a laser slicing process parameter set, and outputting the laser slicing process parameter set to the ultrafast laser control and focusing module and the motion platform and Z-axis closed-loop tracking module. It also receives the wafer quality indicators after stripping from the slicing quality detection and closed-loop feedback unit, and performs process database expansion or neural network incremental learning.

9. The 4H-SiC wafer intelligent laser slicing system as described in claim 8, characterized in that, The ultrafast laser modulation and focusing module includes an ultrafast laser, a half-wave plate, a polarizer, a beam expander, a liquid crystal spatial light modulator, a first lens, a second lens, a dichroic beam splitter, a coaxial CCD camera, and a micro-focusing objective lens equipped with a piezoelectric ceramic axial fine-tuning mechanism. The processing laser output from the ultrafast laser passes sequentially through the half-wave plate and the polarizer to adjust the laser polarization state and output power. The adjusted beam is then incident on the beam expander. The beam expander collimates and magnifies the incident beam, and the processed beam is then incident on the liquid crystal spatial light modulator. The liquid crystal spatial light modulator loads the spherical aberration compensation phase map output from the intelligent process prediction and control system to perform phase correction on the beam to obtain a corrected beam. After phase correction, the corrected beam is output and passes sequentially through the 4F optical relay imaging system composed of the first lens and the second lens. The beam processed by the 4F optical relay imaging system is incident on the dichroic beam splitter. The dichroic beam splitter reflects the corrected beam. The reflected corrected beam is incident on the micro-focusing objective. The micro-focusing objective focuses the corrected beam and projects it into the crystal interior of the 4H-SiC wafer to be processed. The dichroic beam splitter simultaneously transmits the observation light path from the wafer end, and the transmitted observation light path is transmitted to the coaxial CCD camera. The coaxial CCD camera relies on the transmitted light path to complete the coaxial alignment observation of the microscopic features of the wafer surface and the focal length of the spot. The piezoelectric ceramic axial fine adjustment mechanism receives the height control signal output from the motion platform and the Z-axis closed-loop tracking module, and drives the microscope focusing objective to complete the dynamic adjustment of its axial position.

10. The 4H-SiC wafer intelligent laser slicing system as described in claim 8, characterized in that, The motion platform and Z-axis closed-loop tracking module include a vacuum chuck, a horizontal rotation axis, a three-axis linear feed mechanism, and a coaxial laser displacement sensor. The vacuum chuck is used to adsorb and fix the 4H-SiC wafer to be processed, ensuring that the wafer position does not shift during the scanning process. The horizontal rotation axis is used to drive the wafer to rotate and complete the alignment operation between the specific crystal orientation of the wafer and the scanning feed direction. The three-axis linear feed mechanism is used to carry the wafer and complete the high-speed scanning feed and the feed movement in the step direction during the processing. The coaxial laser displacement sensor collects the surface height fluctuation data of the wafer in real time and converts the collected surface height fluctuation data of the wafer into a height signal and outputs it to the intelligent process prediction and control system. The stress-guided peeling unit includes upper and lower opposing tension clamps. The upper and lower opposing tension clamps can apply uniform tensile stress to the 4H-SiC wafer to be processed, which has a micro-explosion modified layer already formed inside, and induce transverse micro-cracks inside the wafer to expand and penetrate along the modified layer, so as to obtain the peeled wafer sheet and realize the peeling of the wafer sheet. The slice quality detection and closed-loop feedback unit is used to measure the surface roughness and critical peeling tensile force of the peeled wafer after peeling is completed, and to obtain the wafer quality index after peeling. When the detection indicators meet the qualification conditions, the slice quality detection and closed-loop feedback unit will send back the relevant data of the qualified sample corresponding to the quality indicators of the peeled wafer to the intelligent process prediction and control system to complete the database expansion. When the detection index exceeds the acceptable range, the slice quality detection and closed-loop feedback unit outputs a recalibration signal to the intelligent process prediction and control system, triggering the online incremental update and adaptive evolution of the neural network.

Images