Semiconductor laser processing defect recognition method and device

Abstract

This invention relates to a method and apparatus for defect identification in semiconductor laser processing. The identification method includes: 1. acquiring laser processing technology and material parameters, and constructing a light-thermal-mechanical conversion evolution model to generate physical state variables of the processing area; 2. acquiring online monitoring data, which includes at least two of coaxial point data, tilt-shift visible light high-speed imaging data, and tilt-shift thermal infrared high-speed imaging data; 3. aligning the online monitoring data with the physical state variables in time and space, and constructing a fusion feature that includes physical model features, online observation features, and predicted observation residual features; 4. inputting the fusion feature into a defect identification model, and outputting the defect category, defect location, severity, and confidence level. This addresses the problems of disconnect between physical mechanisms and online data, lack of a unified alignment mechanism for multimodal data, and insufficient model interpretability and generalization ability. It is applicable to laser processing scenarios for semiconductor wafers, chips, power devices, thin-film devices, and packaging substrates.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor laser processing quality inspection and intelligent control technology, specifically relating to a semiconductor laser processing defect identification method and device driven by the fusion of physical models and online measurement data. This invention is applicable to laser cutting, scribing, micro-hole processing, laser annealing, laser lift-off, and laser micro-welding of semiconductor wafers, chips, power devices, thin-film devices, and packaging substrates. It can accurately identify defects such as microcracks, edge chipping, abnormal recast layers, insufficient ablation, overheating, delamination, abnormal surface roughness, molten pool instability, and abnormal heat-affected zones online. Background Technology

[0002] Semiconductor laser processing is characterized by high energy density, short processing time, large temperature gradient, and fast material response. Processing quality is influenced not only by process parameters such as laser power, scanning speed, spot size, pulse width, defocusing amount, and scanning path, but also by factors such as material absorptivity, thermal conductivity, phase transition characteristics, interlayer bonding in thin films, and the precision of equipment movement. In actual production, defects such as microcracks, edge chipping, incomplete ablation, recast layers, expansion of the heat-affected zone, molten spatter, voids, and delamination often form during millisecond-level thermal cycling and rapid phase transitions. Traditional offline inspection methods, such as offline AOI, microscopy, white light interferometry, or SEM, while highly accurate, suffer from significant hysteresis, making real-time adjustments to the processing impossible and hindering timely identification of the underlying causes of defects.

[0003] Existing online defect identification schemes suffer from three main shortcomings: First, they rely on a single online sensor, such as ordinary camera images or single-point infrared temperature signals. While these schemes can capture some anomalies, they are easily affected by smoke, reflections, depth of focus, noise, and fluctuations in operating conditions, resulting in poor robustness. Second, they directly employ pure data-driven models for defect classification. These schemes are prone to failure when there are insufficient samples, process migrations, material batch changes, or equipment condition fluctuations, and they struggle to explain the physical causal relationships between defects and light absorption, heat accumulation, melt pool evolution, and thermal stress. Third, with the application of coaxial point optical coherence tomography (OCT), high-speed visible light imaging, and high-speed thermal infrared imaging technologies in online monitoring of laser processing, existing technologies typically treat these data as independent observation signals, lacking a unified fusion mechanism with physical evolution models describing the light-heat-force conversion process. This leads to unclear correspondence between monitoring data and defect generation mechanisms, hindering interpretable and highly generalizable defect identification.

[0004] Therefore, how to deeply integrate the physical mechanism of laser processing with multimodal online monitoring data to achieve high-precision, interpretable, and highly generalizable online identification of semiconductor laser processing defects is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies, such as the disconnect between physical mechanisms and online data, the lack of a unified alignment mechanism for multimodal data, and poor model interpretability and generalization ability, and to provide a semiconductor laser processing defect identification method and device driven by the fusion of physical models and online measurement data.

[0006] To achieve the above objectives, the present invention provides the following technical solution.

[0007] A method for identifying defects in semiconductor laser processing includes: S1. Obtain laser processing parameters and material parameters, and construct an optical-thermal-mechanical conversion evolution model based on these parameters to generate physical state variables of the processing area; S2. Acquire online monitoring data, wherein the online monitoring data includes two or three of the following: coaxial point OCT data, tilt-shift visible light high-speed camera data, and tilt-shift thermal infrared high-speed camera data; S3. The online monitoring data and the physical state variables are spatiotemporally aligned, and a fusion feature containing physical model features, online observation features and prediction-observation residual features is constructed; S4. Input the fused features into the defect identification model and output the defect category, defect location, severity, and confidence level.

[0008] Preferably, the light-heat-mechanical conversion evolution model includes: The light energy absorption model is used to generate a space-time heat source term based on laser power, spot distribution, material absorptivity, and surface condition. The heat conduction and phase change model is used to solve the transient temperature field, molten pool or ablation zone morphology of the processing area based on the heat source term and material thermophysical parameters. A thermal stress and defect initiation model is used to calculate thermal stress, strain, and deformation trends based on the temperature field and material mechanical parameters.

[0009] Energy absorption in the light-heat-mechanical conversion evolution model is characterized by the heat source term using the following formula:

[0010] In the formula, η represents the effective absorption coefficient related to wavelength λ, temperature T and surface state s, P(t) represents the instantaneous laser power, g represents the spot energy distribution function, and Veff represents the equivalent action volume. For pulsed lasers, pulse width τ, repetition frequency f and duty cycle parameters are further introduced. For thin film or multilayer semiconductor structures, the absorption coefficient is set as a function of layer thickness and interface reflection.

[0011] The heat conduction and phase transition model obtains the transient temperature field T(x,y,z,t) of the processing area by solving the energy conservation equation, which is:

[0012] In the formula, ρ is the material density, and c p For specific heat capacity, Let k be the temperature gradient, k be the thermal conductivity, Q be the heat source term of the laser volume, L be the latent heat of phase transition, and f be the thermal gradient. s The solid fraction is denoted by ; the boundary conditions of the equations take into account both convective and radiative heat dissipation.

[0013] Preferably, the physical state variables include peak temperature Tmax and temperature gradient. The model considers multiple variables, including cooling rate Rc, heat-affected zone width, aspect ratio of the molten pool or ablation zone, and equivalent stress σeq. When these variables exceed the corresponding material and process window thresholds, the model can generate prior risk assessments for defects such as microcracks, edge chipping, delamination, abnormal recast layer, overheating, or insufficient ablation.

[0014] Preferably, the coaxial point OCT data includes one or more of the following: surface roughness before and after processing, surface height undulation, dynamic three-dimensional morphology of the molten pool during processing, and local recasting accumulation / depression features. The tilt-shift visible light high-speed imaging data includes one or more of the following: surface spatter state, molten pool profile, edge morphology, and material removal dynamics; the tilt-shift thermal infrared high-speed imaging data includes one or more of the following: surface temperature field, heat distribution, hot spot area, heat-affected zone width, and temperature rise / fall rate.

[0015] To effectively correlate multimodal data, the spatiotemporal alignment described in step S3 includes time synchronization and spatial mapping. Time synchronization, based on a unified trigger signal or timestamp, aligns the time of each sensor sequence in the online monitoring data with the processing trajectory. Spatial mapping, based on OCT calibration, camera calibration matrix, laser spot center position, motion trajectory, and workpiece coordinate system, maps the online monitoring data to a unified processing coordinate system, ensuring that OCT measurement points, visible light image pixels, thermal infrared temperature field pixels, and physical state variables correspond to the same processing position or path segment.

[0016] Preferably, the fusion features include physical model output features, online observation features, and prediction-observation residual features. The prediction-observation residual features reflect the degree of deviation between the theoretically stable processing state and the actual measurement, and preferentially include one or more of the following deviations: the deviation between the predicted molten pool scale and the OCT-measured molten pool 3D morphology; the deviation between the predicted molten pool profile and the visible light-measured molten pool profile; and the deviation between the predicted temperature field and the thermal infrared-measured heat distribution.

[0017] Preferably, the defect identification model includes: The OCT morphology coding branch is used to encode the surface morphology and three-dimensional features of the melt pool in OCT data; The visible light temporal coding branch is used to encode the dynamic features of spatter, melt pool profile, and material removal in visible light high-speed camera data; The thermal infrared temporal coding branch is used to encode the temperature field and thermally affected zone evolution characteristics in thermal infrared high-speed camera data; The physical state encoding branch is used to encode the physical state variables; The decision branches are integrated, and the features of each branch are combined and the defect identification results are output through cross-attention, gating fusion, residual fusion, feature splicing or graph structure fusion.

[0018] To ensure the interpretability and physical consistency of the model, the training objectives of the defect identification model include defect classification loss, defect localization loss, severity regression loss, and physical consistency constraint loss. The physical consistency constraint loss is used to constrain the correspondence between the model output and the physical state variables and the online monitoring observation features. The total loss function L can be expressed as: L=Lcls+αLloc+βLsev+γLphy+δLunc In the formula, Lcls represents the defect classification loss, Lloc represents the defect localization loss, Lsev represents the severity regression loss, Lphy represents the physical consistency loss, Lunc represents the uncertainty calibration loss, and α, β, γ, and δ are weighting coefficients. The physical consistency loss is used to constrain the correspondence between the model output and the variables of temperature, stress, cooling rate, OCT morphology change, visible light melt pool profile instability, thermal infrared thermal field anomaly, and residual mutation.

[0019] Preferably, the method further includes step S5: when the severity or confidence level of a defect exceeds a preset threshold, an alarm message, a re-inspection suggestion, or a process parameter adjustment suggestion is generated, and the verified defect sample is fed back into the sample library for iterative updates to the defect identification model. This forms an intelligent closed loop of "monitoring-identification-feedback-optimization," continuously improving processing quality and efficiency.

[0020] The present invention also provides a semiconductor laser processing defect identification device, comprising: The parameter acquisition module is used to acquire laser processing process parameters and material parameters; The photothermal model module is used to construct a photo-thermal-mechanical conversion evolution model and generate physical state variables of the processing area; The online monitoring module includes at least two of the following: a coaxial point OCT module, a tilt-shift visible light high-speed camera module, and a tilt-shift thermal infrared high-speed camera module; The synchronization mapping module is used to align the online monitoring data with the physical state variables in time and space; The feature fusion module is used to construct fused features that include physical model features, online observation features, and prediction-observation residual features; The defect identification module is used to output the defect category, defect location, severity, and confidence level based on the fused features.

[0021] This device can be deployed in the edge computing unit of laser processing equipment, industrial computers, production line quality inspection servers, or cloud-edge collaborative platforms. In one embodiment, the physical structure includes a laser, a beam shaping / collimation unit, a beam scanning / deflection unit, a focusing lens, a semiconductor wafer, and a motion platform; a coaxial point OCT module is arranged coaxially or quasi-coaxially with the processing laser; a tilt-shift visible light high-speed camera module and a thermal infrared high-speed camera module are configured around the processing area; each monitoring module is connected to an industrial computer via a data acquisition and synchronization module, and the defect identification module performs fusion analysis and outputs the results.

[0022] This invention also claims protection for an electronic device, including a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, implements the aforementioned semiconductor laser processing defect identification method.

[0023] And a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described semiconductor laser processing defect identification method.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) High recognition accuracy and strong interpretability. This invention does not simply splice multimodal data, but deeply integrates the light-heat-force conversion evolution model that characterizes the physical mechanism with online monitoring data that reflects on-site disturbances. By constructing a "prediction-observation residual" and introducing physical consistency constraints, the model not only identifies defects, but also correlates them with physical causes such as peak temperature, stress concentration, and splash anomalies, significantly improving the recognition accuracy and the reliability of the results.

[0025] (2) Complete process information and close spatiotemporal correlation. It innovatively combines three highly complementary monitoring methods: coaxial point OCT, tilt-shift visible light high-speed imaging, and tilt-shift thermal infrared high-speed imaging. Through a unified spatiotemporal alignment mechanism, it is correlated to specific processing locations, solving the problem of "difficult alignment and weak correlation" of multi-source heterogeneous data, and providing multi-dimensional information for defect analysis.

[0026] (3) Strong real-time performance, meeting online control requirements. By adopting a reduced-order physical model, sliding window data fusion, and edge inference mechanism, defect risk judgments can be generated in milliseconds or sub-seconds, overcoming the lag of traditional offline detection and making closed-loop real-time control possible.

[0027] (4) Strong generalization ability and adaptability to complex industrial environments. The physical model and physical consistency constraints provide powerful prior knowledge for defect identification, greatly reducing the model's dependence on the number of labeled samples. When material batches are changed, process parameters are fine-tuned, or equipment status drifts, the model, thanks to its physical kernel, exhibits stability and generalization ability far exceeding that of purely data-driven models.

[0028] (5) Strong closed-loop control capability promotes process optimization. The identification results are not only used for alarms, but also associated with specific process parameter adjustment suggestions. Confirmed samples can be fed back into the model for iterative updates, forming an intelligent closed loop of "monitoring-identification-feedback-optimization" to continuously improve processing quality and efficiency. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating the reasoning process of the defect identification method of this invention; Figure 2 This is a schematic diagram of the physical structure of the semiconductor laser processing defect identification device of the present invention; Figure 3 This is a diagram illustrating the structure of the defect identification model training and online inference of this invention. Figure 4 This is a schematic diagram of the photothermal-mechanical conversion evolution and defect formation mechanism in the laser processing process of this invention. Detailed Implementation

[0030] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0031] like Figure 1 As shown, the semiconductor laser processing defect identification method of the present invention includes the following steps: Step S1: Obtain laser processing parameters and material parameters. Process parameters include laser power P, scanning speed v, spot diameter d, wavelength λ, pulse width τ, repetition frequency f, defocusing amount z, scanning path, processing layer number, protective gas flow rate, and ambient temperature. Material parameters include absorptivity, reflectivity, density, specific heat capacity, thermal conductivity, coefficient of thermal expansion, elastic modulus, yield strength, phase transition temperature, latent heat of vaporization, and interlayer bonding parameters. Step S2: Construct a light-heat-force conversion evolution model for the laser processing process and generate physical state variables of the processing area; The model solves or approximates the temperature field, thermal gradient, cooling rate, morphology of the molten or ablated region, thermal stress field and deformation trend of the processing area based on the laser energy input and material response, forming physical state variables that vary with time and space. The light-heat-mechanical conversion evolution model is used to describe the light absorption, heat diffusion, melting or vaporization, solidification and cooling, thermal stress and micro-defect initiation processes that occur in the surface and near-surface regions of a material after laser energy input; the model includes at least a light energy absorption model, a heat conduction and phase transition model, and a thermal stress and defect initiation model. The light energy absorption model generates a space-time heat source term based on laser power, spot distribution, material absorptivity, and surface state. This heat source term can be characterized by the following formula:

[0032] In the formula, η is the effective absorption coefficient related to wavelength λ, temperature T and surface state s, P(t) is the instantaneous laser power, g is the spot energy distribution function, and Veff is the equivalent action volume; for pulsed lasers, pulse width τ, repetition frequency f and duty cycle are further introduced; for thin film or multilayer semiconductor structures, the absorption coefficient is set as a function of layer thickness and interface reflection. The heat conduction and phase change model obtains the transient temperature field T(x,y,z,t) of the processing region by solving the energy conservation equation based on the heat source term and the thermophysical parameters of the material:

[0033] In the formula, ρ is the material density, and c p ρ is the specific heat capacity, k is the thermal conductivity. Let Q be the temperature gradient, Q be the heat source term of the laser volume, L be the latent heat of phase transition, and f be the temperature gradient. s The solid fraction is given; boundary conditions consider both convective and radiative heat dissipation; this model can output the peak temperature Tmax and temperature gradient. Cooling rate Rc, width of heat-affected zone, aspect ratio of molten pool or ablation zone, etc. The thermal stress and defect initiation model calculates thermal stress, strain, and deformation trends based on the temperature field and material mechanical parameters, and outputs equivalent stress σeq, principal tensile stress, stress gradient, and stress concentration factor, etc. When the peak temperature, cooling rate, temperature gradient, or equivalent stress exceeds the material and process window threshold, it generates prior risk of defects such as microcracks, edge chipping, delamination, and recast layer anomalies. Physical state variables include peak temperature Tmax and temperature gradient. Multiple factors, including cooling rate Rc, width of heat-affected zone, aspect ratio of molten pool or ablation zone, and equivalent stress σeq; Step S3: Acquire online monitoring data. The online monitoring data includes two or three of the following: coaxial point OCT data, tilt-shift visible light high-speed camera data, and tilt-shift thermal infrared high-speed camera data. Among them, the coaxial point OCT data includes surface roughness before and after processing, surface height undulation, dynamic three-dimensional morphology of the molten pool during processing, and local recasting accumulation / depression characteristics; the tilt-shift visible light high-speed camera data includes surface spatter state, molten pool contour, edge morphology, and material removal dynamics; the tilt-shift thermal infrared high-speed camera data includes surface temperature field, heat distribution, hot spot area, width of heat-affected zone, and temperature rise / fall rate. Step S4: Spatiotemporal alignment and feature fusion; Spatiotemporal alignment includes time synchronization and spatial mapping; time synchronization is based on a unified trigger signal or timestamp to align each sensor sequence with the processing trajectory in time; spatial mapping is based on OCT calibration, camera calibration matrix, laser spot center position, motion trajectory and workpiece coordinate system to map OCT measurement points, visible light image pixels, thermal infrared temperature field pixels and physical state variables to the same processing position or path segment. The fusion features include at least the physical model output features, online observation features, and prediction-observation residual features. The prediction-observation residual features are used to characterize the deviation between theoretical predictions and actual measurements, and may specifically include: the deviation between the predicted melt pool scale and the OCT measured molten pool 3D morphology; the deviation between the predicted melt pool profile and the visible light measured melt pool profile; and the deviation between the predicted temperature field and the thermal infrared measured heat distribution. Based on a unified trigger signal or timestamp, the OCT measurement sequence, visible light high-speed image sequence, thermal infrared high-speed image sequence, and equipment motion trajectory are mapped to the processing path coordinates; based on the OCT optical axis calibration, camera calibration matrix, laser spot center position, and workpiece coordinate system, the three-dimensional shape points, image pixels, and temperature field regions are associated with the corresponding processing positions; the data is then subjected to denoising, normalization, drift compensation, inter-frame registration, and abnormal segment removal. Step S5: Construct fusion features, which include at least physical model output features, online observation features, and prediction-observation residual features. Among them, online observation features include OCT three-dimensional morphology features, surface roughness features, visible light spatter / contour features, and thermal infrared thermal field features. The residual features are used to characterize the degree of deviation between the theoretical stable processing state and the actual measurement state. Input the fusion features into the defect identification model and output the defect category, defect location, severity, and confidence level.

[0034] like Figure 3 As shown, the defect identification model contains multiple coding branches: The OCT morphology coding branch is used to encode the surface morphology and three-dimensional features of the melt pool in OCT data; The visible light temporal coding branch is used to encode the dynamic features of spatter, melt pool profile, and material removal in visible light high-speed camera data; The thermal infrared temporal coding branch is used to encode the temperature field and thermally affected zone evolution characteristics in thermal infrared high-speed camera data; The physical state coding branch is used to encode physical state variables; The decision branches are integrated by fusing features from each branch through methods such as cross-attention, gating fusion, residual fusion, or feature splicing, and then outputting the defect identification result. During model training, the objective function includes defect classification loss, defect localization loss, severity regression loss, and physical consistency constraint loss. The physical consistency constraint loss is used to constrain the correspondence between the model output and variables such as temperature, stress, cooling rate, OCT morphology changes, visible light melt pool profile instability, thermal infrared thermal field anomalies, and residual mutations. The total loss function can be expressed as: L=Lcls+αLloc+βLsev+γLphy+δLunc In the formula, α, β, γ, and δ are weighting coefficients; Step S6: Train the accurate defect identification model. The model includes an OCT morphology encoding branch, a visible light temporal encoding branch, a thermal infrared temporal encoding branch, a physical state encoding branch, and a fusion decision branch. The training objectives include defect classification loss, defect localization loss, severity regression loss, and physical consistency constraint loss. When the severity or confidence of a defect exceeds a preset threshold, alarm information, re-inspection suggestions, or process parameter adjustment suggestions are generated. The verified defect samples are then fed back into the sample library for iterative updates to the defect identification model.

[0035] Step S7: Online identification and closed-loop feedback. During the processing, the model outputs the defect category, location, severity level and confidence level in real time. When the defect risk exceeds the set threshold, the device outputs an alarm, a re-inspection suggestion or a process parameter adjustment suggestion, and feeds the confirmed samples back into the sample library for model iteration.

[0036] The models are described in detail below: 1. Photothermal-mechanical conversion evolution model: The photo-thermal-mechanical conversion evolution model is used to describe the processes of light absorption, thermal diffusion, melting or vaporization, solidification and cooling, thermal stress, and micro-defect initiation that occur in the surface and near-surface regions of a material after laser energy input. This model can be implemented using one or more of the following methods: finite element method, finite difference method, finite volume method, analytical approximation, reduced-order model, or physically constrained neural network.

[0037] (1) Light energy absorption model: In the processing coordinate system, the laser energy input is represented as a space-time heat source term related to power, spot size, wavelength, pulse width, scanning speed, and material surface state. For example, the heat source term can be represented as: Q(x,y,z,t)=η(λ,T,s)·P(t)·g(x-x0(t),y-y0(t),z,d) / Veff In the formula, η represents the effective absorption coefficient related to wavelength λ, temperature T, and surface state s, P(t) represents the instantaneous laser power, g represents the spot energy distribution function, and Veff represents the equivalent action volume. For pulsed lasers, pulse width τ, repetition frequency f, and duty cycle parameters can be further introduced; for thin film or multilayer semiconductor structures, the absorption coefficient can be set as a function of layer thickness and interface reflection.

[0038] (2) Heat conduction and phase transition model: The heat conduction process can be described by the energy conservation equation. Considering factors such as material density, specific heat, thermal conductivity, latent heat of phase change, convective heat dissipation, and radiative heat dissipation, the transient temperature field of the processing area can be obtained. The model output should include at least the peak temperature Tmax and the temperature gradient. Indicators such as cooling rate Rc, width of heat-affected zone, and width-to-depth ratio of molten pool or ablation zone.

[0039] For rapid online applications, high-precision finite element models can be solved offline in advance to form parameterized response surfaces or reduced-order models. In the online stage, the temperature field, molten pool morphology and defect risks can be quickly inferred based on real-time process parameters and OCT / visible light / thermal infrared measurement feedback, thereby meeting the requirements for millisecond or subsecond defect identification.

[0040] (3) Thermal stress and defect initiation model: Based on the temperature field, thermal stress, strain, and deformation trends are further calculated according to thermal expansion, elastic or elastoplastic constitutive relations, interlaminar constraints, and cooling contraction effects. The output of the thermal stress model includes equivalent stress, principal tensile stress, stress gradient, residual stress risk, and stress concentration factor.

[0041] When the peak temperature, cooling rate, temperature gradient, stress concentration, or molten pool morphology index exceeds the corresponding material and process window threshold, the model can generate prior knowledge of defects such as microcracks, edge chipping, delamination, abnormal recast layer, overheating, or insufficient ablation, providing physically interpretable input for the subsequent fusion identification model.

[0042] Table 1. Correspondence between physical variables and defects in the thermal stress and defect initiation model

[0043] 2. Online monitoring data fusion: The online monitoring data in this invention is not simply spliced ​​together, but rather fused using a unified time base, unified spatial coordinates, and unified process semantics. Preferably, the online monitoring data primarily consists of coaxial point OCT data, tilt-shift visible light high-speed camera data, and tilt-shift thermal infrared high-speed camera data, and can be combined with process parameters and equipment feedback data. Specifically, the processing includes the following steps.

[0044] Table 2. Data Categories and Main Functions in Online Monitoring Data Fusion

[0045] (1) Time synchronization: Each acquisition module receives a unified trigger signal from the laser controller or motion controller, or is aligned using an industrial clock, timestamp, and sampling frequency conversion. For modules with different sampling frequencies, a sliding time window is used for aggregation, for example, mapping OCT point measurement sequences, visible light high-speed image frames, thermal infrared high-speed image frames, and equipment motion trajectories to the same processing path point or the same processing sub-region.

[0046] (2) Spatial mapping: Based on OCT optical path calibration, visible light camera calibration matrix, thermal infrared camera calibration matrix, laser spot center position, motion trajectory, and workpiece coordinate system, OCT 3D topography points, visible light image pixels, and thermal infrared temperature field pixels are converted into processing coordinates. For curve scanning, array processing, and multilayer thin film structures, spatial indexes can be established according to path segments, processing units, layer numbers, or device units to bind defect locations with process parameters, physical state variables, and online observation features.

[0047] (3) Construction of fusion features: The fusion features include at least three categories: The first category is physical model features, such as peak temperature, thermal gradient, cooling rate, molten pool aspect ratio, heat-affected zone width, and equivalent stress; the second category is online observation features, including surface roughness, surface height undulation, dynamic 3D morphology of the molten pool, and local accumulation / depression features from coaxial point OCT, the number of spatters, spatter direction, molten pool profile size, edge integrity, and surface anomaly dynamics from tilt-shift visible high-speed imaging, and the highest temperature, temperature distribution uniformity, hot spot area, heat-affected zone width, and temperature rise / fall rate from tilt-shift thermal infrared high-speed imaging; the third category is physical-observation residual features, including the deviation between the predicted molten pool scale and the OCT 3D morphology measurement results, the deviation between the predicted molten pool profile and the visible light observation profile, and the deviation between the predicted temperature field and the thermal infrared observation heat distribution.

[0048] 3. Precise Defect Identification Model: The defect identification model is used to output defect identification results based on fused features. The model can employ deep neural networks, gradient boosting trees, support vector machines, Bayesian models, physically constrained neural networks, or combinations thereof; preferably, the model adopts a multimodal coding and physically constrained fusion structure.

[0049] (1) Model structure: The model includes an OCT morphology encoding branch, a visible light temporal encoding branch, a thermal infrared temporal encoding branch, a physical state encoding branch, a fusion decision branch, and an output branch. The OCT morphology encoding branch encodes features such as surface roughness, surface height undulation, and dynamic 3D morphology of the molten pool; the visible light temporal encoding branch encodes spatter, molten pool contour, edge integrity, and material removal dynamics; the thermal infrared temporal encoding branch encodes temperature field, hot spots, heat-affected zone, and temperature rise / fall process; the physical state encoding branch encodes variables such as temperature field, thermal gradient, cooling rate, stress field, and molten pool geometry; and the fusion decision branch learns the contribution weights of different data sources at different defect formation stages through cross-attention, gated fusion, feature stitching, residual fusion, or graph neural networks.

[0050] (2) Training objectives: During the training phase, offline detection labels or manually verified labels are used as supervision signals, while physical consistency constraints are introduced. The total loss function can be expressed as: L=Lcls +αLloc+βLsev+γLphy+δLunc In the formula, Lcls represents the defect classification loss, Lloc represents the defect localization loss, Lsev represents the severity regression loss, Lphy represents the physical consistency loss, Lunc represents the uncertainty calibration loss, and α, β, γ, and δ are weighting coefficients. The physical consistency loss can be used to constrain the correspondence between the model output and variables such as temperature, stress, cooling rate, OCT morphology changes, visible light melt pool profile instability, thermal infrared thermal field anomalies, and residual mutations.

[0051] (3) Output results: Model outputs may include: defect type, such as microcracks, chipping, burrs, insufficient ablation, overheating, recast layer anomalies, spatter, delamination, abnormal surface roughness, and abnormal heat-affected zone; defect location, such as path coordinates, wafer coordinates, chip cell number, or image region; severity level, such as minor, moderate, severe, or scored continuously from 0 to 1; confidence level and physical explanation, such as being mainly caused by excessively high peak temperature, excessive cooling rate, abnormal OCT 3D morphology, instability of visible light molten pool profile, or abnormal thermal infrared heat distribution.

[0052] Figure 2The physical structure of the semiconductor laser processing defect identification device of the present invention is shown. The device includes a laser, a beam shaping / collimation unit, a beam scanning / deflection unit, a focusing lens, a semiconductor wafer / workpiece, and a motion platform; a coaxial point OCT module is arranged coaxially or quasi-coaxially with the processing laser to collect surface roughness before and after processing and the dynamic three-dimensional morphology of the molten pool during processing; a shift-axis visible light high-speed camera module and a shift-axis thermal infrared high-speed camera module are configured around the processing area to observe surface spatter and molten pool contours, as well as the evolution of the temperature field and heat-affected zone, respectively; each module is connected to an industrial computer via a data acquisition and synchronization module, where the photothermal model module, synchronization mapping module, feature fusion module, and defect identification module deployed on the computer perform analysis, and the defect identification results are output through a display / interactive module.

[0053] This device can be deployed in the edge computing unit of laser processing equipment, industrial computers, production line quality inspection servers, or cloud-edge collaboration platforms. The device includes the following modules: In this embodiment, as Figure 2 As shown, the physical structure of the device includes a laser, a beam shaping / collimation unit, a beam scanning / deflection unit, a focusing lens, a semiconductor wafer / workpiece, and an XYθ motion platform; a point-type OCT module, an axis-shifting visible light high-speed camera module, and an axis-shifting thermal infrared high-speed camera module are arranged coaxially or quasi-coaxially with the processing laser around the processing area; the above monitoring modules are connected to an industrial computer / controller via a data acquisition and synchronization module, and the defect identification module performs fusion analysis and outputs the defect identification results through a display / interactive module.

[0054] Preferably, the coaxial point OCT module is used to acquire surface roughness before and after processing, as well as the dynamic three-dimensional morphology of the molten pool during processing; the tilt-shift visible light high-speed camera module is used to acquire the spatter state, molten pool contour, and dynamic surface morphology of the processed surface; and the tilt-shift thermal infrared high-speed camera module is used to acquire information on heat distribution, temperature evolution, and heat-affected zone in the processed area. In other embodiments, photoelectric, acoustic emission, vibration, or other auxiliary sensor data can also be extended, but this does not affect the technical solution of the present invention, which mainly uses OCT, visible light, and thermal infrared data for fusion recognition.

[0055] (1) Parameter acquisition module, used to acquire laser processing process parameters, material parameters, equipment status parameters and processing path information.

[0056] (2) Online monitoring module, used to collect coaxial point OCT data, tilt-shift visible light high-speed camera data, tilt-shift thermal infrared high-speed camera data, and can selectively collect equipment feedback and other auxiliary sensor data.

[0057] (3) Synchronization mapping module, used to map online monitoring data to unified processing spatiotemporal coordinates based on unified timestamp, trigger signal, OCT calibration, camera calibration matrix and motion trajectory.

[0058] (4) Photothermal model module, used to generate physical state variables such as temperature field, thermal gradient, cooling rate, molten pool morphology, thermal stress and deformation risk based on process parameters and material parameters.

[0059] (5) Feature fusion module, used to construct physical model features, OCT morphology features, visible light dynamic features, thermal infrared thermal field features and prediction-observation residual features, and perform normalization, dimensionality reduction, attention weighting or gating fusion.

[0060] (6) Defect identification module, used to output defect category, defect location, severity level and confidence level based on fused features.

[0061] (7) Early warning feedback module, used to generate alarm information, re-inspection suggestions, process parameter adjustment suggestions and sample refill information based on defect identification results.

[0062] (8) Model update module, used to add samples that have been detected offline or verified manually to the sample library, and to periodically or incrementally update the model thresholds or parameters.

[0063] The following specific examples further illustrate this point.

[0064] Example 1: Defect Identification in Wafer Laser Dicing Process Step S1: Parameter Acquisition During the wafer laser dicing process, the device acquires the wafer dicing process parameters and material parameters. The process parameters include: laser power P = 5W, scanning speed v = 200mm / s, spot diameter d = 20μm, wavelength λ = 355nm, pulse width τ = 10ns, repetition frequency f = 100kHz, and defocusing amount z = 0. Material parameters include silicon's absorptivity, density, specific heat capacity, thermal conductivity, coefficient of thermal expansion, elastic modulus, and phase transition temperature.

[0065] Step S2: Constructing a light-heat-mechanical conversion evolution model Based on the above parameters, a photothermal-mechanical transformation evolution model is constructed, including a laser energy absorption model, a heat conduction and phase transition model, and a thermal stress and defect initiation model. A reduced-order model (pre-calculated offline finite element response surface) is used to achieve rapid online inference. The model outputs physical state variables, including the peak temperature T_max and temperature gradient along the dicing path. Cooling rate R_ c. Width of the heat-affected zone and equivalent stress σ _eq; Step S3: Obtain online monitoring data During the zoning process, three types of online monitoring data were collected simultaneously: Coaxial point optical coherence tomography (OCT) data: Surface roughness (Ra=0.1μm) and height undulation near the dicing path before and after processing were measured, and the dynamic three-dimensional morphology of the molten pool was acquired during processing. The width of the molten pool was recorded as approximately 22μm and the depth as approximately 25μm. Tilt-shift visible light high-speed camera data: Images of surface spatter, molten pool profile, and edge integrity were acquired at a rate of 10,000 frames per second. The images show a low number of spatters, a regular molten pool profile, and no edge chipping. Tilt-shift thermal infrared high-speed camera data: Temperature field and heat-affected zone evolution were acquired at a rate of 2,000 frames per second. Thermal images show a peak temperature of approximately 1,200°C, a heat-affected zone width of approximately 40 μm, and uniform temperature distribution. Step S4: Spatiotemporal alignment and data preprocessing Based on a unified hardware trigger signal, the timestamps of all data are aligned. According to OCT optical axis calibration, camera calibration matrix, and motion trajectory coordinates, the OCT 3D topography points, visible light image pixels, and thermal infrared temperature field pixels are uniformly mapped to the world coordinate system of the wafer dicing, corresponding to specific dicing path points (e.g., the 5th dicing path, 10mm from the starting point), forming a multimodal data packet. The data is then filtered, denoised, and normalized. Step S5: Construct fusion features For each aligned spatiotemporal location, construct a fused feature vector: Physical model characteristics: T_max = 1, 180°C =8×10 6 K / m, R_c=1.2×10 6 K / s, σ_eq=150MPa.

[0066] Online observation characteristics: OCT surface roughness Ra=0.12μm, molten pool width 22μm; visible light spatter count=2, molten pool profile roundness=0.95; thermal infrared peak temperature 1,200°C, heat-affected zone width 40μm.

[0067] Prediction-observation residual characteristics: The deviation between the predicted melt pool width (model prediction of 20 μm) and the OCT measured width (22 μm) is +2 μm; the deviation between the predicted peak temperature (model prediction of 1,160°C) and the thermal infrared observed peak temperature (1,200°C) is +40°C. Step S6: Defect Identification The aforementioned fused features are input into a pre-trained defect recognition model. This model includes an OCT shape encoding branch (one-dimensional convolutional network), a visible light temporal encoding branch (Transformer encoder), a thermal infrared temporal encoding branch (long short-term memory network), a physical state encoding branch (fully connected network), and a fusion decision branch based on cross-attention. Model output results: The defect category is "no defect" or "slight surface roughness anomaly", the defect location is the path point coordinates (x=125.3mm, y=45.8mm), the severity score is 0.3 (0-1 scale), and the confidence level is 0.92; Step S7: Closed-loop feedback (optional) Since the severity of the defect did not exceed the preset threshold (0.7), the system only recorded the result and did not trigger an alarm. Subsequent offline testing confirmed that there was no functional defect at this location, and this sample can be used as a positive sample to be fed back into the sample library for subsequent incremental learning of the model.

[0068] Example 2: Laser Grooving Defect Identification in Silicon Carbide Devices This embodiment is basically the same as Embodiment 1, except that the object being processed is a silicon carbide device, and the processing technology is laser grooving. Silicon carbide material is hard and brittle with high thermal conductivity, making it extremely prone to cracking and edge chipping.

[0069] During the grooving process, the optical-thermal-mechanical model predicted that the stress concentration factor of a certain groove segment suddenly increased to 2.5 (threshold 1.8); at the same time, the high-speed visible light image shifted by the axis showed that the spatter in this groove segment increased significantly, and discontinuities appeared in the groove opening profile; the three-dimensional morphology of the coaxial point OCT showed that there was a local abrupt change in depth at the bottom of the groove (the depth changed from the expected 50 μm to 65 μm) and sidewall protrusions; the thermal infrared image showed that the temperature distribution in this area was uneven, and local hot spots appeared. The defect identification model integrates the aforementioned physical residuals (stress concentration factor exceeding the threshold) and multimodal observation anomalies (increased spatter, discontinuous contours, abrupt changes in three-dimensional morphology), outputting "edge chipping" and "microcrack" defects with high confidence (0.98). The defect location is accurate to the specific groove segment coordinates (e.g., the 3rd grooving line, 15.2 mm from the starting point), and the severity level is "severe." The early warning feedback module immediately triggers an alarm and suggests reducing the laser power by 10% or increasing the scanning speed by 15%. This result is sent to the equipment controller to achieve real-time adjustment of process parameters.

[0070] Example 3: Layer Identification During Thin Film Laser Lifting This embodiment is basically the same as Embodiment 1, except that the processing technology is thin film laser peeling.

[0071] The device acquires parameters such as film thickness, interlayer absorptivity, and interfacial bonding strength, which are input into a photo-thermal-mechanical model to predict interfacial temperature and thermal stress distribution. During online monitoring, high-speed thermal infrared imaging revealed abnormal local heat accumulation in a certain area, with the temperature exceeding the model's prediction by 80°C. Coaxial point OCT measurements showed that the surface roughness Ra of this area after peeling was 0.25 μm, significantly higher than the 0.08 μm of the normal area. Visible light images showed that the peeling front stagnated and expanded non-uniformly in this area. After comprehensive analysis, the defect identification model outputs "uneven delamination" and "local over-peeling" defects with high confidence, and provides a physical explanation: mainly caused by abnormal local heat accumulation leading to excessively high interface temperature and thermal stress exceeding the bonding strength. The system suggests adjusting the scanning speed or using a multi-pass scanning strategy to improve uniformity.

[0072] Creative explanation of the overall technical solution: The technical solution claimed in this invention has outstanding substantive features and significant progress compared with the prior art, and the specific inventive step is described as follows: This paradigm innovation differs from existing technologies in that it integrates physical models with data-driven approaches. Existing technologies are mainly divided into three categories: purely offline detection (lagging), single / multi-sensor data-driven (lacking mechanism, prone to failure), and purely physical simulation (computationally demanding, unable to reflect real-time disturbances). This invention breaks away from these frameworks, pioneering a deep fusion paradigm that uses an optical-thermal-mechanical conversion evolution model as an interpretable prior and multimodal online sensing data as the source of real-time correction and residuals. Its key innovation lies in: This invention is not a simple stitching process: it does not use the physical model as an offline analysis tool, nor does it simply stitch together optical coherence tomography, visible light, and thermal infrared data into a multi-channel image. Instead, it constructs a closed-loop logic of "physical prediction → online observation → residual calculation → fusion discrimination." Among these features, the "prediction-observation residual characteristic" is one of the key technical features that distinguishes this invention from all existing technologies. This residual characteristic directly quantifies the deviation between the theoretically stable processing state and the actual field state, serving as a sensitive indicator for triggering defect early warning and locating the cause of defects.

[0073] Deep fusion of spatiotemporal semantics: Existing technologies often process sensor data independently and without alignment. This invention unifies heterogeneous data with different physical meanings (3D morphology, 2D images, temperature fields), different sampling frequencies, and different coordinate systems into the same spatiotemporal coordinates of the processing path through a unified time synchronization and spatial mapping mechanism. This provides an aligned input with clear process semantics for subsequent fusion recognition models, representing a substantial improvement in the field of multimodal information processing.

[0074] Key means to solve the technical problem of "mechanism-data disconnect": The core technical problem this invention aims to solve is the "disconnect between physical mechanisms and online data." To address this, this invention provides a complete and closed-loop technical approach: Constructing a computable physical bridge: The specific components of the light-heat-mechanical conversion evolution model (energy absorption, heat conduction phase transition, thermal stress) and their resulting physical state variables (peak temperature, stress, etc.) are not dispensable computational byproducts, but rather serve as "hidden states" or "bottleneck features" connecting processing parameters (cause) and final defects (effect). They provide a physical inductive bias for the data-driven model, enabling it to learn rapidly even with limited samples.

[0075] Physical Consistency Constraint: This invention introduces a "physical consistency constraint loss," a crucial step in "embedding" physical mechanisms into the deep learning model training process. It not only requires the model's judgment results to be accurate but also demands that the model's internal feature representations conform to physical laws (e.g., the model's output crack risk should be positively correlated with the predicted thermal stress magnitude and consistent with the observed abrupt changes in cooling rate). This significantly suppresses the "shortcut learning" and "spurious correlation" problems of purely data-driven models, improving the model's robustness and interpretability.

[0076] Targeted selection and system integration of multi-source heterogeneous sensor data: This invention does not refer to "multi-sensor fusion" in general, but specifically defines a particular sensor combination of coaxial point optical coherence tomography, tilt-shift visible light high-speed imaging, and tilt-shift thermal infrared high-speed imaging, based on the characteristics of defects in semiconductor laser processing.

[0077] Synergy: These three aspects are observed from three completely orthogonal physical dimensions: "three-dimensional micro-morphology" (molten pool morphology in the initiation stage, roughness after solidification), "two-dimensional macro-dynamics" (splashes, contours), and "thermodynamic state" (temperature field, heat-affected zone), forming a seamless perception coverage of the light-heat-mechanical coupling process. All three are indispensable; any combination of two cannot achieve the complete observation effect of this invention.

[0078] Targeted approach: This combination is designed to address the formation mechanisms of semiconductor processing-specific defects such as microcracks, edge chipping, recast layers, and delamination. For example, without optical coherence tomography (OCT), it is impossible to accurately characterize micron-scale recast layers and subsurface damage; without thermal infrared spectroscopy, it is impossible to quantify the heat-affected zone and cooling rate, key factors that generate thermal stress.

[0079] Significant technological progress A balance between real-time performance and accuracy: Existing technologies either suffer from poor real-time performance (offline detection) or poor accuracy and stability (single sensor). This invention, through a reduced-order physical model and a multimodal fusion network, achieves for the first time, millisecond-level online response and high-precision (near-offline detection) defect identification within the same framework, representing a significant performance leap.

[0080] Explainable intelligence: Traditional deep learning models are "black boxes." The output of this invention not only tells "what the defect is," but also explains "why" through residual features and physical variables (such as "thermal stress cracks caused by excessively rapid cooling rates"), which has immeasurable value for production sites that require rapid process intervention.

[0081] Cross-condition generalization capability: When material batches change, the performance of purely data-driven models drops sharply. However, the physical model part of this invention can readjust its output according to the new material parameters, providing a correct baseline for the identification model, thereby significantly improving the system's adaptability to process changes.

[0082] In summary, this invention organically combines a specific physical model, a specific sensor combination, and an innovative spatiotemporal-residual fusion framework to form a complete and highly effective technical solution that differs from any existing technology. This solution is not a simple combination of existing technical features, but rather a novel paradigm proposed to solve the recognized technical challenge of "precise online monitoring of laser processing." The various technical features support and synergize with each other, producing a remarkable technical effect of "1+1>2," demonstrating significant inventiveness.

[0083] The above description is only a preferred embodiment of the present invention. All equivalent changes and modifications made within the scope of the claims of the present invention should be covered by the claims of the present invention.

Claims

1. A method for identifying defects in semiconductor laser processing, characterized by, include: S1. Obtain laser processing parameters and material parameters, and construct an optical-thermal-mechanical conversion evolution model based on these parameters to generate physical state variables of the processing area; S2. Acquire online monitoring data, wherein the online monitoring data includes two or three of the following: coaxial point data, tilt-shift visible light high-speed camera data, and tilt-shift thermal infrared high-speed camera data; S3. The online monitoring data and the physical state variables are spatiotemporally aligned, and a fusion feature containing physical model features, online observation features and prediction-observation residual features is constructed; S4. Input the fused features into the defect identification model and output the defect category, defect location, severity, and confidence level.

2. The method of claim 1, wherein the step of identifying the defect comprises the steps of: determining a location of the defect; and determining a type of the defect. The light-heat-mechanical conversion evolution model includes: The light energy absorption model is used to generate a space-time heat source term based on laser power, spot distribution, material absorptivity, and surface condition. The heat conduction and phase change model is used to solve the transient temperature field, molten pool or ablation zone morphology of the processing area based on the heat source term and material thermophysical parameters. A thermal stress and defect initiation model is used to calculate thermal stress, strain, and deformation trends based on the temperature field and material mechanical parameters. Energy absorption in the light-heat-mechanical conversion evolution model is characterized by the heat source term using the following formula: Q(x,y,z,t)=η(λ,T,s)·P(t)·g(x-x0(t),y-y0(t),z,d) / Veff In the formula, η represents the effective absorption coefficient related to wavelength λ, temperature T and surface state s, P(t) represents the instantaneous laser power, g represents the spot energy distribution function, and Veff represents the equivalent action volume; for pulsed lasers, pulse width τ, repetition frequency f and duty cycle parameters are further introduced; for thin film or multilayer semiconductor structures, the absorption coefficient is set as a function of layer thickness and interface reflection. The heat conduction and phase transition model obtains the transient temperature field T(x,y,z,t) of the processing area by solving the energy conservation equation, which is: where p is the material density, c p is the specific heat capacity, k is the thermal conductivity, is the temperature gradient, Q is the laser body heat source term, L is the latent heat of phase change, f s is the solid fraction; the boundary conditions of the equation take into account the convective and radiative heat dissipation.

3. The method according to any one of claims 1 to 2, wherein The physical state variables include a peak temperature Tmax, a temperature gradient , a cooling rate Rc, a heat-affected zone width, a molten pool or ablation region width-to-depth ratio, and an equivalent stress σeq.

4. The semiconductor laser processing defect identification method according to claim 1, characterized in that, Coaxial point OCT data includes one or more of the following: surface roughness before and after machining, surface height undulation, dynamic three-dimensional morphology of the molten pool during machining, and local recasting accumulation / depression features; The tilt-shift visible high-speed camera data includes one or more of the following: surface spatter state, molten pool profile, edge morphology, and material removal dynamics; the tilt-shift thermal infrared high-speed camera data includes one or more of the following: surface temperature field, heat distribution, hot spot area, heat-affected zone width, and temperature rise / fall rate.

5. The method of claim 1, wherein the step of identifying the defect comprises the step of: The spatiotemporal alignment mentioned in step S3 includes: ​ Time synchronization: Based on a unified trigger signal or timestamp, the time of each sensor sequence in the online monitoring data is aligned with the processing trajectory. Spatial mapping: Based on OCT calibration, camera calibration matrix, laser spot center position, motion trajectory and workpiece coordinate system, the online monitoring data is mapped to a unified processing coordinate system, so that the OCT measurement point, visible light image pixel, thermal infrared temperature field pixel and the physical state variable correspond to the same processing position or path segment.

6. The method of claim 1, wherein the step of identifying the defect comprises the step of: The prediction-observation residual features include one or more of the following deviations: the deviation between the predicted melt pool scale and the OCT measured three-dimensional morphology of the melt pool; the deviation between the predicted melt pool profile and the visible light measured melt pool profile; and the deviation between the predicted temperature field and the thermal infrared measured heat distribution. ​ 7. The method of claim 1, wherein the step of identifying the defect comprises the steps of: determining a location of the defect; and determining a type of the defect. The defect identification model includes: The OCT morphology coding branch is used to encode the surface morphology and three-dimensional features of the melt pool in OCT data; The visible light temporal coding branch is used to encode the dynamic features of spatter, melt pool profile, and material removal in visible light high-speed camera data; The thermal infrared temporal coding branch is used to encode the temperature field and thermally affected zone evolution characteristics in thermal infrared high-speed camera data; The physical state encoding branch is used to encode the physical state variables; The decision branches are integrated, and the features of each branch are combined and the defect identification results are output through cross-attention, gating fusion, residual fusion, feature splicing or graph structure fusion.

8. The method according to claim 1 or 7, wherein The training objectives of the defect identification model include defect classification loss, defect localization loss, severity regression loss, and physical consistency constraint loss. The physical consistency constraint loss is used to constrain the correspondence between the model output and the physical state variables and online monitoring observation features. The total loss function L for training the defect identification model is expressed as: L=Lcls+αLloc+βLsev+γLphy+δLunc In the formula, Lcls represents the defect classification loss, Lloc represents the defect localization loss, Lsev represents the severity regression loss, Lphy represents the physical consistency loss, Lunc represents the uncertainty calibration loss, and α, β, γ, and δ are weighting coefficients. The physical consistency loss is used to constrain the correspondence between the model output and the variables of temperature, stress, cooling rate, OCT morphology change, visible light melt pool profile instability, thermal infrared thermal field anomaly, and residual mutation.

9. The method of claim 1, wherein the step of identifying the defect comprises the step of: Also includes: ​ S5. When the severity or confidence level of the defect exceeds a preset threshold, generate alarm information, re-inspection suggestions or process parameter adjustment suggestions, and feed the verified defect samples back into the sample library for iterative updates to the defect identification model.

10. A semiconductor laser processing defect identification device, characterized in that, include: The parameter acquisition module is used to acquire laser processing process parameters and material parameters; The photothermal model module is used to construct a photo-thermal-mechanical conversion evolution model and generate physical state variables of the processing area; The online monitoring module includes at least two of the following: a coaxial point OCT module, a tilt-shift visible light high-speed camera module, and a tilt-shift thermal infrared high-speed camera module; The synchronization mapping module is used to align the online monitoring data with the physical state variables in time and space; The feature fusion module is used to construct fused features that include physical model features, online observation features, and prediction-observation residual features; The defect identification module is used to output the defect category, defect location, severity, and confidence level based on the fused features.

11. An electronic device comprising a processor and a memory, the memory having stored therein a computer program, characterized in that, When the computer program is executed by the processor, it implements the semiconductor laser processing defect identification method of claim 1.

12. A computer readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the processor, it implements the semiconductor laser processing defect identification method of claim 1.

Images