A method and system for rapid detection of surface defects of silicon carbide MOS devices

Abstract

The application discloses a kind of silicon carbide MOS device surface defect rapid detection method and system, it is related to the technical field of non-pattern wafer defect detection.A kind of silicon carbide MOS device surface defect rapid detection system, comprising have: risk distribution identification module, scanning path generation module, detection data acquisition module, defect detection module and defect analysis module.The application is introduced by defect risk assessment model, utilizes historical detection data and sample detection data to generate risk distribution heat map and detection round information, realizes the accurate calibration of device risk area and the dynamic control of detection round;By constructing historical detection channel and sample detection channel, long-term risk indicators and fitting type regional risk indicators are extracted respectively, and the two are fused into comprehensive risk probability distribution, taking into account batch statistical law and current sample characteristics.

Description

Technical Field

[0001] This invention relates to the field of patternless wafer defect detection technology, and in particular to a rapid detection method and system for surface defects in silicon carbide MOS devices. Background Technology

[0002] Silicon carbide MOSFETs, with their excellent high-temperature, high-voltage, and high-frequency performance, have become core power devices in key fields such as new energy vehicles, rail transportation, and smart grids. As manufacturing precision continues to improve and end-user applications demand higher reliability, the impact of minute defects on device surface on overall performance and yield is becoming increasingly significant. Especially in large-scale production, failure to identify high-risk defect areas in a timely manner can easily lead to batch failures, resulting in significant quality risks and economic losses.

[0003] Existing defect detection methods mostly rely on full-coverage scanning or image recognition, which suffers from low efficiency, limited data dimensions, and poor ability to identify weak scattering signals, making it difficult to optimize detection efficiency while ensuring accuracy. Furthermore, most methods lack the ability to model the correlation between historical defect distribution patterns and current device sample characteristics, making it difficult to dynamically adjust detection strategies to adapt to devices with different batches and structural differences. This also hinders the achievement of a high degree of matching between the detection path and risk areas, further restricting the improvement of intelligent inspection levels on production lines.

[0004] Developing an efficient, accurate, and dynamically adaptable defect detection method that can achieve rapid and selective scanning detection of high-risk areas on the surface of silicon carbide MOS devices while ensuring the accuracy of defect detection, thereby improving detection efficiency and supporting the dynamic detection needs in large-scale production, has become a key technological requirement in the silicon carbide device industry chain. Summary of the Invention

[0005] This invention proposes a rapid detection method based on the collaborative drive of risk assessment and dual-channel defect identification model. By introducing a detection round control strategy, a partitioned interval detection mechanism, and risk area marking output, it achieves dual assurance of efficiency and accuracy in defect detection.

[0006] A rapid detection method for surface defects in silicon carbide MOS devices includes:

[0007] Historical testing data of devices or sample testing data are acquired and input into the defect risk assessment model to obtain testing round information and risk distribution heatmap; the risk distribution heatmap includes risk areas and corresponding risk levels; the testing round information includes round parameters and round update parameters, and the round parameters are used to divide consecutive devices into testing round groups;

[0008] Based on round parameters and risk distribution heatmap, an adaptive scanning path is generated; the adaptive scanning path executes a partitioned interval detection strategy for risk areas in the detection round group that are below the set risk level.

[0009] The device to be inspected is positioned on a motion platform. The platform's movement and laser irradiation are controlled according to an adaptive scanning path to excite and receive scattering signals related to defects, thereby acquiring scattering data for a single device. The scattering data of all devices in each inspection round are selectively merged into joint inspection data.

[0010] The scattering data and joint detection data are preprocessed and feature extracted to obtain multidimensional surface defect data; the multidimensional surface defect data are then input into the defect identification model for processing to obtain defect detection results with risk area markings.

[0011] The defect density and distribution of the defect detection results are analyzed, each device is comprehensively scored to obtain the defect level, the corresponding defect spatial distribution map is output, and the detection round information and risk distribution heatmap are updated.

[0012] As a preferred technical solution of the present invention, the defect risk assessment model is constructed based on a gradient boosting decision tree, and supervised training is performed using labeled samples containing historical detection data and sample detection data to construct historical detection channels and sample detection channels respectively.

[0013] As a preferred technical solution of the present invention, the historical detection channel statistically analyzes the real defect records of multiple batches of devices through historical detection data, extracts the defect occurrence frequency, stability and evolution trend with process changes in each spatial region, and constructs a long-term risk index; the sample detection channel analyzes the structural characteristics, material consistency and response signal characteristics of each region in the sample detection data of the current device, and compares and fits them with the regional risk characteristics of similar devices in pre-training, to generate a fitted regional risk index for the current device; the risk indices output by the two channels are combined through a fusion function to generate a comprehensive risk probability distribution of the current device in each spatial region, and each region is marked as a risk region of different risk levels according to the set division criteria, which is used to construct a risk distribution heatmap, and combined with a preset round control strategy to generate detection round information.

[0014] As a preferred embodiment of the present invention, the round parameters are used to set corresponding partitioned interval detection strategies for the risk areas corresponding to each risk level; the partitioned interval detection strategy is used to perform full coverage sampling of the target risk area within the detection round group according to a preset spatial interval rule, ensuring that at least full coverage sampling of each risk area is completed within the detection round group; the round update parameters are used to control the update speed of the round parameters.

[0015] As a preferred embodiment of the present invention, the generation of the adaptive scanning path includes: setting scanning priorities for risk areas corresponding to different risk levels based on round parameters and the distribution of risk areas in the risk distribution heatmap; generating corresponding scanning coverage density parameters according to the priority of each risk area and the partitioned interval detection strategy, and planning the platform movement path and the laser irradiation position; the platform movement path is a linear or spiral scanning trajectory defined based on the structural layout of the device, and its step spacing is adjusted according to the risk level of the corresponding risk area; the laser irradiation position is planned according to a preset incident angle combination and detection point sequence; finally, an adaptive scanning path that meets the round detection requirements and the risk distribution coverage balance is generated.

[0016] As a preferred technical solution of the present invention, the selective merging of scattering data from all devices in each detection round group into joint detection data includes: identifying scattering data from each device that has completed full coverage sampling of the target risk area based on a partitioned interval detection strategy; performing spatial correspondence and quality screening on scattering data from risk areas of the same risk level that have undergone full coverage sampling, and constructing joint detection data.

[0017] As a preferred embodiment of the present invention, the acquisition of multidimensional surface defect data includes: performing signal preprocessing operations on the scattering data of a single device, including wavelet denoising, normalization, and spectral enhancement processing on the original scattering signal, and extracting independent defect data containing spatial angle, intensity frequency, and time series features; reconstructing and aligning the joint detection data according to the spatial division of risk areas in the partitioned interval detection strategy, and extracting joint defect data with stability features, background response contours, and similarity features using feature mapping and density consistency analysis; merging the independent defect data and joint defect data into multidimensional surface defect data, and inputting it into a dual-channel defect recognition model for processing.

[0018] As a preferred embodiment of the present invention, the defect identification model includes a dual-channel structure, which receives independent defect data and joint defect data from multidimensional surface defect data as inputs respectively. The dual-channel structure extracts the spatial frequency features and local response patterns of the two types of inputs through a parallel convolutional neural network, and performs cross-channel fusion using an inter-channel feature matching mechanism. The fusion method includes weight reconstruction based on an attention mechanism and a residual connection structure. The fused features are used to determine defects through a fully connected classifier, and the output includes defect detection results including defect location, defect type, and risk area markings corresponding to the risk level.

[0019] As a preferred technical solution of the present invention, the method of obtaining the defect level includes: statistically analyzing the risk area markings in the defect detection results to obtain the distribution density and aggregation degree of defects in each region of the device surface; comprehensively scoring each device based on the spatial diffusion range, density threshold and risk level ratio of the defect location; constructing the defect level based on the comprehensive scoring results and generating a corresponding spatial distribution map to show the spatial location, risk intensity and distribution pattern of defects on the device surface for subsequent process screening, quality control and production feedback.

[0020] A rapid detection system for surface defects in silicon carbide MOS devices includes:

[0021] Risk distribution identification module: acquires historical testing data of devices or sample testing data, inputs it into the defect risk assessment model, and obtains testing round information and risk distribution heatmap;

[0022] Scan path generation module: Generates adaptive scan paths based on round parameters and risk distribution heatmaps;

[0023] Detection data acquisition module: Based on the adaptive scanning path control platform movement and laser irradiation, it acquires the scattering data of individual devices; and selectively merges the scattering data of all devices in each detection round group into joint detection data;

[0024] Defect detection module: preprocesses and extracts features from scattering data and joint detection data respectively to obtain multidimensional surface defect data; inputs the multidimensional surface defect data into the defect recognition model for processing to obtain defect detection results with risk area markings;

[0025] Defect Analysis Module: Analyzes the defect density and distribution of defect detection results, performs a comprehensive score for each device to obtain the defect level, outputs the corresponding defect spatial distribution map, and updates the detection round information and risk distribution heatmap.

[0026] The present invention has the following advantages:

[0027] This invention introduces a defect risk assessment model, using historical and sample testing data to generate risk distribution heatmaps and testing round information. This enables precise calibration of device risk areas and dynamic control of testing rounds, thereby improving the targeting and adaptability of the testing process. By constructing historical and sample testing channels, long-term risk indicators and fitted regional risk indicators are extracted respectively, and the two are integrated into a comprehensive risk probability distribution, taking into account both batch statistical patterns and current sample characteristics, significantly improving the accuracy and universality of risk assessment.

[0028] This invention achieves a balance between detection coverage and detection efficiency by setting round parameters and round update parameters, and by adopting a partitioned interval full coverage sampling strategy within the detection round group. This avoids over-detection of low-risk areas and omission of high-risk areas. By planning an adaptive scanning path based on the risk distribution heatmap and round parameters, and by comprehensively optimizing scanning priority, scanning coverage density parameters, platform motion path, and laser irradiation position, the invention achieves adaptive adjustment of the detection path, thereby improving scanning speed and detection accuracy.

[0029] This invention improves the utilization rate and information redundancy of multi-device data by spatially corresponding and quality screening scattering data of the same risk level within the detection round group and constructing joint detection data; by combining the independent defect data of a single device with the joint defect data of multiple devices to form multidimensional surface defect data, it captures both local features and overall distribution features, ensuring the comprehensiveness and multidimensional expression of defect information.

[0030] This invention employs a dual-channel defect identification model to extract spatial frequency features and local response patterns from independent and joint defect data in parallel. It also enhances the model's ability to identify complex defects through a cross-channel fusion mechanism, achieving high-precision output of defect location, defect type, and risk area marking.

[0031] This invention performs statistical analysis on the risk area markings in the defect detection results and constructs a defect spatial distribution map to intuitively display the distribution density, diffusion range and aggregation characteristics of defects, thereby providing a quantitative basis for process screening, quality control and production feedback. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are only schematic diagrams of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0033] Figure 1 This is a schematic diagram of a rapid detection system for surface defects of silicon carbide MOS devices used in an embodiment of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0035] In embodiments of the present invention, "surface defects" refer to microscopic discontinuities distributed on or near the surface of a silicon carbide MOS device that may affect the device's electrical performance and structural reliability. These defects include, but are not limited to: surface microcracks, trench defects caused by uneven etching, foreign bumps caused by particles or residues, abnormal oxide layer thickness or uniformity, excessively rough step edges, mismatched regions extending to the surface due to lattice dislocations, and other abnormal response regions exhibiting abnormal intensity, angular distortion, or spectral distortion under optical or scattering signals. These defects can be isolated points or exhibit banded, clustered, or diffused distribution characteristics, all falling within the scope of "surface defects" as described in this specification.

[0036] Example 1: A rapid detection method for surface defects in silicon carbide MOS devices, comprising the following steps:

[0037] Step S1: Obtain historical testing data of the device or sample testing data, input them into the defect risk assessment model, and obtain testing round information and risk distribution heatmap; the risk distribution heatmap includes risk areas and corresponding risk levels; the testing round information includes round parameters and round update parameters, the round parameters are used to divide consecutive devices into testing round groups;

[0038] The defect risk assessment model is constructed based on a gradient boosting decision tree and uses labeled samples containing historical detection data and sample detection data for supervised training to construct historical detection channels and sample detection channels respectively.

[0039] The historical inspection channel uses historical inspection data to statistically analyze the actual defect records of multiple batches of devices, extracts the frequency of defect occurrence, stability, and evolution trend with process changes in each spatial region, and constructs long-term risk indicators. The sample inspection channel analyzes the structural characteristics, material consistency, and response signal characteristics of each region in the sample inspection data of the current device, and compares and fits them with the regional risk characteristics of similar devices that have been pre-trained, generating a fitted regional risk indicator for the current device. The risk indicators output by the two channels are combined through a fusion function to generate a comprehensive risk probability distribution of the current device in each spatial region. According to the set division criteria, each region is marked as a risk region of different risk levels to construct a risk distribution heatmap, and combined with a preset round control strategy to generate inspection round information.

[0040] The round parameters are used to set corresponding partitioned interval detection strategies for the risk areas corresponding to each risk level. The partitioned interval detection strategy is used to perform full coverage sampling of the target risk area within the detection round group according to the preset spatial interval rules, ensuring that at least full coverage sampling of each risk area is completed within the detection round group. The round update parameters are used to control the update speed of the round parameters.

[0041] In this embodiment, historical inspection data is derived from complete inspection records of multiple batches of silicon carbide MOS devices on the production line. This dataset includes: device batch number, device serial number, inspection timestamp, inspection area coordinates or serial number, defect type, defect location distribution, defect size information, and confirmation markers. Based on this, the statistically derived long-term risk index reflects the frequency of defect occurrence in a specific spatial region across multiple batches. For example, if microcrack defects appear in ten consecutive batches in the gate edge region, this region is marked as a high-risk area. Simultaneously, stability is characterized by calculating the standard deviation of the defect occurrence probability in the same region across different batches. If the frequency of a certain defect continuously increases under different process parameters, the evolution trend is recorded as an upward risk index, used to indicate that subsequent inspection strategies need to increase the inspection density in this region.

[0042] The sample testing data originates from the initial characterization phase of the device under test, primarily including regional structural features (such as trench depth and step edge geometry), material consistency indicators (such as oxide layer thickness uniformity and local doping concentration deviation), and response signal characteristics (such as optical scattering intensity distribution and spectral response characteristics). For example, when an oxide layer thickness deviation exceeding a preset threshold is detected in a localized region, that region is assigned a higher risk score. The sample testing channel compares these features with the regional risk features already labeled in the similar device database. If the response characteristics of the current device are highly similar to those of historically high-risk devices in a specific region, the regional risk index obtained through fitting will be significantly improved.

[0043] The risk distribution heatmap is generated based on the device surface coordinates, marking each region according to different risk levels (high, medium, and low) and attaching a risk probability value. For example, in a certain region at the source edge of a power device, if the historical defect frequency exceeds 30% and the current sample detection results are highly consistent with high-risk characteristics, then this region is presented as a red high-risk area in the heatmap.

[0044] Round control strategy is used to transform risk distribution heatmaps into actionable inspection round information. Round parameters are used to divide continuously inspected devices into inspection round groups, for example, every 20 devices are grouped together. Simultaneously, inspection strategies are set according to different risk levels; for example, high-risk areas must be fully covered in every round group, medium-risk areas must be covered at least once every two round groups, and low-risk areas must be sampled and covered every three round groups. Round update parameters dynamically evaluate the rationality of the inspection strategy by recording the defect detection rate and the number of defect-free periods for each area in multiple rounds of inspection. For example, if a low-risk area has no defects for three consecutive rounds, the round update parameters will reduce its subsequent inspection frequency; if a medium-risk area has a defect in a certain round, it will be immediately adjusted to a high-risk area, and the inspection density will be increased accordingly.

[0045] The training process of the defect risk assessment model includes: using historical detection data and sample detection data as input features, and defect occurrence status as a supervision label, iterative training is performed using a gradient boosting decision tree algorithm. The model learns the nonlinear mapping relationship between different features and defect risk. During model operation, the historical detection channel and the sample detection channel output regional risk indicators respectively. After being weighted and combined by a fusion function, a comprehensive risk probability distribution is obtained. Then, risk level labels are generated by thresholding, and finally, a risk distribution heatmap and detection round information are output.

[0046] Step S2: Based on the round parameters and risk distribution heatmap, generate an adaptive scanning path; the adaptive scanning path executes a partitioned interval detection strategy for risk areas in the detection round group that are below the set risk level;

[0047] The adaptive scanning path generation includes: setting scanning priorities for risk areas corresponding to different risk levels based on round parameters and the distribution of risk areas in the risk distribution heatmap; generating corresponding scanning coverage density parameters according to the priority of each risk area and the partitioned interval detection strategy, and planning the platform movement path and laser irradiation position; the platform movement path is a linear or spiral scanning trajectory defined based on the device's structural layout, and its step spacing is adjusted according to the risk level of the corresponding risk area. The higher the risk level, the smaller the scanning spacing and the higher the coverage accuracy; the laser irradiation position is planned according to a preset combination of incident angles and detection point sequences, so that during the platform movement, multi-angle dense irradiation of high-risk areas and skip sampling of low-risk areas can be achieved; finally, an adaptive scanning path that meets the round detection requirements and risk distribution coverage balance is generated.

[0048] The adaptive scan path has the ability to dynamically respond to round-up parameter updates. Specifically, after the detection of a detection round group is completed, the round parameters will be adjusted in real time based on the detection results of different risk level areas within that round group, thereby dynamically modifying the scan path of subsequent products.

[0049] The round update parameters reflect the statistical trend of defects appearing or remaining defect-free in multiple rounds of detection for each risk level area, and are used to evaluate the rationality of the current detection strategy and the redundancy of risk control. For example, if no defects are found in a low-risk area in several consecutive rounds, the detection frequency of that area in subsequent rounds is reduced based on a preset adjustment threshold, which is manifested by increasing the spacing of sampling points in that area in the scanning path or skipping steps; while if an area was originally designated as a medium-risk area, and an anomaly is detected in a round, the risk level of that area is increased, and subsequent scanning paths are modified accordingly to configure denser scanning coverage in that area.

[0050] By linking the cycle update parameters with the planning of adaptive scanning paths, the scanning paths are no longer fixed trajectories but have the ability to evolve in real time. The detection strategy is adjusted based on the feedback information of the current batch of products, thereby improving detection efficiency and resource utilization while ensuring defect coverage.

[0051] The partitioned interval detection strategy in this embodiment refers to setting unequal sampling intervals in the spatial dimension based on the risk level of the risk area. High-risk areas adopt dense scanning, that is, each sub-area is fully covered and sampled; medium-risk areas adopt interval scanning, that is, some sub-areas are covered according to preset intervals to ensure representative sampling within a round group; low-risk areas adopt skip sampling, only some key points are detected to reduce unnecessary resource consumption.

[0052] The scanning coverage density parameters include: step spacing, with micrometer-level step spacing for high-risk areas, sub-millimeter-level step spacing for medium-risk areas, and further relaxation for low-risk areas; coverage factor, with multi-angle coverage required for high-risk areas, single-angle coverage sufficient for medium-risk areas, and fewer coverage times for low-risk areas; and angle combination, with full-angle incidence required for high-risk areas, and only representative angles used for low-risk areas.

[0053] The platform's motion path is generated based on the device layout coordinates. When the device has a rectangular layout, a linear scan trajectory is preferred; when the device is circular or polygonal, a spiral trajectory is preferred. The scan trajectory planning must ensure priority coverage of high-risk areas. For example, in high-risk sections at the edge of the device gate, the platform path will automatically compress the step spacing to form high-density coverage.

[0054] Parametric description of laser irradiation position: The laser irradiation position consists of three parts, including the combination of incident angles, the sequence of detection points, and the corresponding timing control. High-risk areas are assigned multiple incident angles (such as 0°, 30°, 60°) and scanned repeatedly in each detection point sequence; low-risk areas use a single incident angle and the number of sampling points is reduced to reduce detection redundancy.

[0055] Once a detection round is completed, the defect detection results are immediately fed back to the risk distribution heatmap and round update parameters. The scanning path is automatically adjusted based on the updated parameters; for example, in the path generation for the next round, the scanning density of a certain area may be directly increased or decreased. This dynamic response mechanism avoids the resource waste associated with fixed paths and enables the detection strategy to learn and evolve.

[0056] For example, if a region that was originally marked as low risk is found to have microcrack defects in the first round of inspection, then when generating the scanning path for the next round, the region will be upgraded to a medium-risk region, and the scanning interval will be reduced and the angle coverage increased accordingly to ensure the accuracy of subsequent inspections.

[0057] Step S3: Position the device to be inspected on the motion platform, control the platform movement and laser irradiation according to the adaptive scanning path, excite and receive the scattering signal related to the defect, and obtain the scattering data of a single device; selectively merge the scattering data of all devices in each inspection round group into joint inspection data;

[0058] The selective merging of scattering data from all devices in each detection round group into joint detection data includes: identifying scattering data from each device that has completed full coverage sampling of the target risk area based on a partitioned interval detection strategy; performing spatial correspondence and quality screening on scattering data from risk areas of the same risk level that have undergone full coverage sampling, and constructing joint detection data.

[0059] In this embodiment, scattering data refers to the raw record of optical scattering signals caused by defects on the device surface and its surrounding structures under laser irradiation, including multi-dimensional information such as time-domain signal waveform, frequency-domain amplitude spectrum, scattering angle distribution, and energy intensity. Scattering data acquisition is completed by the linkage between the motion platform and the laser irradiation device. The laser is incident on different positions on the device surface sequentially according to an adaptive scanning path, and the photodetector receives the reflected and scattered signals at the corresponding angles and stores them digitally at a high sampling rate.

[0060] The motion platform possesses high-precision displacement control capabilities, employing sub-micron stepper motors and a position feedback system to ensure accurate positioning of the device in both the two-dimensional plane and the height direction. The device is fixed to the platform via vacuum adsorption or mechanical clamps to prevent displacement or tilting during high-speed scanning, ensuring the stability of scattering data acquisition.

[0061] For each device, scattering data is collected by region, with each risk region corresponding to a scattering signal sequence. High-risk regions collect data with higher dimensionality, including multiple incident angles, multiple detection angles, and repeated sampling sequences; low-risk regions only collect basic angle signals. For example, in high-risk regions, each detection point can acquire scattering signals at three incident angles: 0°, 30°, and 60°, with at least two repeated acquisitions at each angle for subsequent denoising and stability analysis.

[0062] After a detection round is completed, the scattering data from all devices need to be selectively merged. The merging rules include:

[0063] Coverage filtering: Only scattering data that meet the full coverage sampling conditions are included in the merging, that is, ensuring that the area has reached the preset sampling density in both spatial and angular dimensions.

[0064] Spatial alignment: By mapping device layout coordinates, data from the same region in different devices are spatially aligned. For example, scattering data from the same region located at the gate edge in all devices are uniformly mapped to a standard coordinate grid.

[0065] Quality screening: The signal-to-noise ratio, stability, and drift of the scattered signals are evaluated, outliers and low-quality data are removed, and only data that meets the threshold requirements are retained for merging.

[0066] After the above processing, the joint detection data includes: a list of source devices and batch information; regional identification and risk level; an aligned scattering signal matrix (aligned in time domain, frequency domain, and angular dimensions); quality weight information (used to characterize the contribution of different source data to the merging result); and regional statistical characteristics, such as mean response curve, spectral main peak distribution, and angular energy distribution.

[0067] For example, in a certain round, high-risk areas of all ten devices underwent multi-angle complete coverage sampling. After spatial correspondence and quality screening, joint detection data for that area was obtained. The merged result includes not only the aligned data matrix of the scattered signals from the ten devices, but also the mean scattering spectrum and standard deviation curve of the area, which are used to reflect the commonalities and differences in the defect responses of the area.

[0068] After the joint detection data is constructed, it is indexed with the risk distribution heatmap and round parameters for subsequent model input and updates. If a certain area shows abnormal fluctuations in the merged results, it is marked as an area requiring key monitoring in the round update parameters, and its sampling density is increased in the next round of scanning.

[0069] Step S4: Preprocess and extract features from the scattering data and joint detection data respectively to obtain multidimensional surface defect data; input the multidimensional surface defect data into the defect recognition model for processing to obtain defect detection results with risk area markings;

[0070] The acquisition of multidimensional surface defect data includes: performing signal preprocessing operations on the scattering data of individual devices, including wavelet denoising, normalization, and spectral enhancement processing on the original scattering signals, and extracting independent defect data containing spatial angle, intensity frequency, and time series features; reconstructing and aligning the joint detection data according to the spatial division of risk areas in the partitioned interval detection strategy, and extracting joint defect data with stability features, background response contours, and similarity features using feature mapping and density consistency analysis; merging the independent defect data and joint defect data into multidimensional surface defect data, and inputting it into a dual-channel defect recognition model for processing.

[0071] The defect identification model includes a dual-channel structure, receiving independent defect data and joint defect data from multidimensional surface defect data as inputs. The independent defect data characterizes the defect response features of the current product, while the joint defect data provides background response reference and feature compensation information for the target area. The dual-channel structure extracts the spatial frequency features and local response patterns of the two types of inputs through a parallel convolutional neural network, and performs cross-channel fusion using an inter-channel feature matching mechanism. The fusion method includes weight reconstruction based on an attention mechanism and a residual connection structure. The fused features are then used by a fully connected classifier to determine defects, outputting defect detection results including defect location, defect type, and risk area markings corresponding to the risk level.

[0072] In this embodiment, signal preprocessing includes: wavelet denoising, which uses a multi-scale wavelet decomposition method to remove high-frequency noise from the scattered signal while preserving local features caused by defects; normalization, which adjusts the scattering intensity at different detection points to a uniform scale, eliminating the impact of light source power fluctuations and detector gain differences between devices on the data; and spectrum enhancement, which uses Fourier transform and spectral smoothing techniques to enhance the response characteristics of the defect signal within a specific frequency range, improving the ability to detect weak defects.

[0073] Independent defect data comes from the scattering signals of a single device and, after preprocessing, includes: spatial angular characteristics (scattering distribution under different incident angles); intensity-frequency characteristics (distribution curve of scattered energy in the frequency domain and main peak frequency); and time series characteristics (dynamic response changes over time).

[0074] For example, in a high-risk area, independent defect data manifests as a shift in the main peak of the spectrum and abrupt changes in the amplitude of the time-domain signal. These characteristics directly indicate potential microcracks or foreign matter residues.

[0075] The joint defect data is obtained by merging the scattering signals of multiple devices in the same risk region and includes: stability features: a consistency index of the scattering response of different devices in this region; background response profile: the average scattering characteristic curve of the defect-free sample, used as a comparison baseline; similarity features: a measure of the difference between the current device and the joint sample, used to identify local anomalies.

[0076] For example, when joint data shows that a certain region responds stably in most devices, but deviates significantly in the current device, that region is identified as a high-risk defect point.

[0077] The formation of multidimensional surface defect data: Independent defect data and joint defect data are aligned in the same coordinate system and feature dimensions to form multidimensional surface defect data. This data structure has high dimensionality and multi-channel features, which can simultaneously reflect the defect characteristics of individual devices and the statistical characteristics of group devices, thereby ensuring the accuracy and robustness of defect identification.

[0078] Defect identification models include:

[0079] Input layer: Independent defect data and joint defect data are fed into two convolutional neural network channels respectively;

[0080] Convolutional extraction: Spatial frequency patterns and local response patterns are extracted independently for each channel, and high-risk defects can produce significant feature differences at this stage;

[0081] Cross-channel fusion: The importance weights of each feature are calculated through an attention mechanism, and the complementarity between channels is enhanced by a residual connection structure;

[0082] Classifier output: The fully connected classifier classifies the fused features and outputs the defect location, defect type, and risk area markers corresponding to the risk level.

[0083] Risk region marking not only indicates the location coordinates of defects but also includes the risk level and type of the defect. For example, when a microcrack is detected in a source edge region, the model output will mark this region as "high-risk - microcrack," and use it as the basis for constructing the defect spatial distribution map in subsequent steps. For instance, in a single detection, independent defect data may show a shift in the main peak of the scattered signal spectrum in a certain region, while joint defect data may show a stable background response in the same region. After the model fusion is performed, it is determined that this region is significantly different from the overall trend, and the output defect type is "surface particle residue," the risk level is "medium risk," and this is recorded in the risk region marking.

[0084] Step S5: Analyze the defect density and distribution of the defect detection results, perform a comprehensive score for each device to obtain the defect level, output the corresponding defect spatial distribution map, and update the detection round information and risk distribution heatmap.

[0085] The process of obtaining the defect level includes: statistically analyzing the risk area markers in the defect detection results to obtain the distribution density and aggregation degree of defects in each area of ​​the device surface; comprehensively scoring each device based on the spatial diffusion range, density threshold, and risk level ratio of the defect location; constructing the defect level based on the comprehensive scoring results and generating a corresponding spatial distribution map to show the spatial location, risk intensity, and distribution pattern of defects on the device surface for subsequent process screening, quality control, and production feedback.

[0086] In this embodiment, the defect detection results include the defect location, defect type, and risk area markers corresponding to the risk level. During analysis, the device surface is first divided into regions according to the spatial division of the risk distribution heatmap. The number and area of ​​defects within each region are calculated as the defect density, and the concentration or dispersion of defects in spatial distribution is further evaluated. For example, when the number of defects in a certain region accounts for more than 20% of the region's area and is concentrated in a local lattice, it is defined as a high-density region.

[0087] Spatial diffusion range describes the ductility of defects on the device surface. If a defect spans multiple regions and its total coverage area exceeds a preset threshold, it is considered a diffuse defect. The density threshold is set based on statistical experience to distinguish between normal process fluctuations and abnormal defect distributions. For example, when the defect density exceeds 10⁻³ / μm², it is determined to be an out-of-range defect distribution.

[0088] In the comprehensive scoring process, defects of different risk levels are assigned different weights, with high-risk defects having a significantly higher weight than medium- and low-risk defects. For example, the weight of a high-risk defect is set to 1.0, medium-risk to 0.6, and low-risk to 0.3. The final comprehensive score is obtained through weighted calculation to ensure that the impact of high-risk defects on the scoring results is fully amplified.

[0089] The overall scoring results are divided into several grade ranges to determine the overall defect level of the device. The defect level system provides a clear grading basis for subsequent production screening.

[0090] Level 1: No obvious defects or only low-risk defects exist, and the overall score is lower than the set threshold A;

[0091] Level 2: There are a few medium-risk defects, and the overall score is between threshold A and threshold B;

[0092] Level 3: There are obvious high-risk defects or large-area defect clusters, and the comprehensive score exceeds the threshold B.

[0093] The defect spatial distribution map uses the device surface coordinates as a reference to visualize the defect detection results as a two-dimensional layer. The map displays the defect location, defect type, and corresponding risk level. For example, high-risk defects are marked with red areas, medium-risk defects with orange areas, low-risk defects with green areas, and areas with high defect density are displayed with darker colors.

[0094] After the defect level and defect spatial distribution map are generated, the risk distribution heatmap and inspection round information are updated. If a certain region repeatedly shows high-risk defects in multiple devices, the risk level of that region in the risk heatmap is increased. At the same time, the round update parameters record the defect trend of that region, and the sampling frequency of that region is increased in the adaptive scan path of subsequent round groups. Conversely, if a low-risk region remains defect-free for a long period, its coverage density in the scan path is reduced to improve overall inspection efficiency.

[0095] For example, in one inspection, microcracks repeatedly appeared in the source edge region of a batch of devices, causing the defect density in that area to be far higher than the set threshold. The final comprehensive score exceeded threshold B, and the devices were rated as level three defects. In the corresponding defect spatial distribution map, this area was marked as a large red high-risk zone. Simultaneously, the round update parameters marked this area as persistently high-risk. In the next inspection round, the scanning interval for this area was further reduced, and multi-angle illumination was increased to ensure inspection coverage.

[0096] Example 2: A rapid detection system for surface defects in silicon carbide MOS devices, see [link to example]. Figure 1 As shown, it includes the following modules:

[0097] Risk distribution identification module: acquires historical testing data of devices or sample testing data, inputs it into the defect risk assessment model, and obtains testing round information and risk distribution heatmap;

[0098] Scan path generation module: Generates adaptive scan paths based on round parameters and risk distribution heatmaps;

[0099] Detection data acquisition module: Based on the adaptive scanning path control platform movement and laser irradiation, it acquires the scattering data of individual devices; and selectively merges the scattering data of all devices in each detection round group into joint detection data;

[0100] Defect detection module: preprocesses and extracts features from scattering data and joint detection data respectively to obtain multidimensional surface defect data; inputs the multidimensional surface defect data into the defect recognition model for processing to obtain defect detection results with risk area markings;

[0101] Defect Analysis Module: Analyzes the defect density and distribution of defect detection results, performs a comprehensive score for each device to obtain the defect level, outputs the corresponding defect spatial distribution map, and updates the detection round information and risk distribution heatmap.

[0102] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. 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 rapid detection method for surface defects in silicon carbide MOS devices, characterized in that, include: Historical testing data of the device or sample testing data are obtained and input into the defect risk assessment model to obtain testing round information and risk distribution heatmap; The risk distribution heatmap includes risk areas and their corresponding risk levels; The detection round information includes round parameters and round update parameters. The round parameters are used to divide consecutive devices into detection round groups. Based on round parameters and risk distribution heatmap, an adaptive scanning path is generated; the adaptive scanning path executes a partitioned interval detection strategy for risk areas in the detection round group that are below the set risk level. The device to be inspected is positioned on a motion platform. The platform's movement and laser irradiation are controlled according to an adaptive scanning path to excite and receive scattering signals related to defects, thereby acquiring scattering data for a single device. The scattering data of all devices in each inspection round are selectively merged into joint inspection data. The scattering data and joint detection data are preprocessed and feature extracted respectively to obtain multidimensional surface defect data; Multidimensional surface defect data is input into a defect identification model for processing to obtain defect detection results with risk area markings; The defect density and distribution of the defect detection results are analyzed, each device is comprehensively scored to obtain the defect level, the corresponding defect spatial distribution map is output, and the detection round information and risk distribution heatmap are updated.

2. The rapid detection method for surface defects in a silicon carbide MOS device according to claim 1, characterized in that, The defect risk assessment model is constructed based on a gradient boosting decision tree and uses labeled samples containing historical detection data and sample detection data for supervised training to construct historical detection channels and sample detection channels respectively.

3. The rapid detection method for surface defects in a silicon carbide MOS device according to claim 2, characterized in that, The historical inspection channel uses historical inspection data to statistically analyze the actual defect records of multiple batches of devices, extracts the frequency of defect occurrence, stability, and evolution trend with process changes in each spatial region, and constructs long-term risk indicators. The sample detection channel analyzes the structural features, material consistency, and response signal characteristics of each region in the sample detection data of the current device, and compares and fits them with the regional risk characteristics of similar devices that have been pre-trained, generating a fitted regional risk index for the current device. The risk indices output by the two channels are combined through a fusion function to generate a comprehensive risk probability distribution of the current device in each spatial region. Based on the set division criteria, each region is marked as a risk region of different risk levels to construct a risk distribution heatmap, and detection round information is generated in combination with a preset round control strategy.

4. The method for rapid detection of surface defects in a silicon carbide MOS device according to claim 1, characterized in that, The round parameters are used to set corresponding partitioned interval detection strategies for the risk areas corresponding to each risk level. The partitioned interval detection strategy is used to perform full coverage sampling of the target risk area within the detection round group according to the preset spatial interval rules, ensuring that at least full coverage sampling of each risk area is completed within the detection round group; The round update parameter is used to control the update speed of the round parameter.

5. The rapid detection method for surface defects in a silicon carbide MOS device according to claim 1, characterized in that, The adaptive scanning path generation includes: setting scanning priorities for risk areas corresponding to different risk levels based on round parameters and the distribution of risk areas in the risk distribution heatmap; generating corresponding scanning coverage density parameters according to the priority of each risk area and the partitioned interval detection strategy, and planning the platform movement path and the laser irradiation position; the platform movement path is a linear or spiral scanning trajectory defined based on the structural layout of the device, and its step spacing is adjusted according to the risk level of the corresponding risk area; the laser irradiation position is planned according to a preset incident angle combination and detection point sequence; finally, an adaptive scanning path that meets the round detection requirements and risk distribution coverage balance is generated.

6. The method for rapid detection of surface defects in a silicon carbide MOS device according to claim 1, characterized in that, The selective merging of scattering data from all devices in each detection round group into joint detection data includes: identifying scattering data from each device that has completed full coverage sampling of the target risk area based on a partitioned interval detection strategy; performing spatial correspondence and quality screening on scattering data from risk areas of the same risk level that have undergone full coverage sampling, and constructing joint detection data.

7. The rapid detection method for surface defects in a silicon carbide MOS device according to claim 1, characterized in that, The acquisition of multidimensional surface defect data includes: performing signal preprocessing operations on the scattering data of individual devices, including wavelet denoising, normalization, and spectral enhancement processing on the original scattering signals, and extracting independent defect data containing spatial angle, intensity frequency, and time series features; reconstructing and aligning the joint detection data according to the spatial division of risk areas in the partitioned interval detection strategy, and extracting joint defect data with stability features, background response contours, and similarity features using feature mapping and density consistency analysis; merging the independent defect data and joint defect data into multidimensional surface defect data, and inputting it into a dual-channel defect recognition model for processing.

8. The rapid detection method for surface defects in a silicon carbide MOS device according to claim 1, characterized in that, The defect identification model includes a dual-channel structure, which receives independent defect data and joint defect data from multidimensional surface defect data as inputs. The dual-channel structure extracts the spatial frequency features and local response patterns of the two types of inputs through a parallel convolutional neural network, and performs cross-channel fusion using an inter-channel feature matching mechanism. The fusion method includes weight reconstruction based on an attention mechanism and a residual connection structure. The fused features are used to determine defects through a fully connected classifier, and the output includes defect detection results including defect location, defect type, and risk area markings of corresponding risk levels.

9. The rapid detection method for surface defects in a silicon carbide MOS device according to claim 1, characterized in that, The process of obtaining the defect level includes: statistically analyzing the risk area markers in the defect detection results to obtain the distribution density and aggregation degree of defects in each area of ​​the device surface; comprehensively scoring each device based on the spatial diffusion range, density threshold, and risk level ratio of the defect location; constructing the defect level based on the comprehensive scoring results and generating a corresponding spatial distribution map to show the spatial location, risk intensity, and distribution pattern of defects on the device surface for subsequent process screening, quality control, and production feedback.

10. A rapid detection system for surface defects in silicon carbide MOS devices, characterized in that, The system employs a rapid surface defect detection method for silicon carbide MOS devices according to any one of claims 1 to 9, comprising: Risk distribution identification module: acquires historical testing data of devices or sample testing data, inputs it into the defect risk assessment model, and obtains testing round information and risk distribution heatmap; Scan path generation module: Generates adaptive scan paths based on round parameters and risk distribution heatmaps; Detection data acquisition module: Based on the adaptive scanning path control platform movement and laser irradiation, it acquires the scattering data of individual devices; and selectively merges the scattering data of all devices in each detection round group into joint detection data; Defect detection module: preprocesses and extracts features from scattering data and joint detection data respectively to obtain multidimensional surface defect data; inputs the multidimensional surface defect data into the defect recognition model for processing to obtain defect detection results with risk area markings; Defect Analysis Module: Analyzes the defect density and distribution of defect detection results, performs a comprehensive score for each device to obtain the defect level, outputs the corresponding defect spatial distribution map, and updates the detection round information and risk distribution heatmap.

Images