Silicon carbide epitaxial layer morphology defect classification method and system based on reinforcement learning
By using a reinforcement learning-based method for classifying defects in silicon carbide epitaxial layers and dynamically adjusting the probe beam attitude, the problem of distinguishing between artifacts and deep dislocations in existing technologies is solved. This achieves efficient defect identification and prediction, improving production yield and device reliability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SUZHOU MACROCORE SEMICON CO LTD
- Filing Date
- 2026-04-20
- Publication Date
- 2026-05-22
AI Technical Summary
Existing technologies struggle to accurately distinguish between normal growth artifacts and real defects caused by step flow at a 4° off-axis angle under dynamic thermal stress conditions, and cannot effectively differentiate between surface contamination and deep dislocations, resulting in a high false positive rate and inaccurate defect identification.
A reinforcement learning-based method for classifying morphological defects in silicon carbide epitaxial layers is adopted. By acquiring historical morphological evolution data and multi-view morphological response tensors, a multi-level spatial domain analysis subsystem and a reinforcement learning agent are constructed to dynamically adjust the attitude of the probe beam, thereby distinguishing between step-gathering artifacts and real defects.
It reduces the false positive rate, improves the accuracy and reliability of defect identification, provides a basis for predicting whether defects will expand into fatal failure sources, and improves production yield and device reliability.
Smart Images
Figure CN122072691A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of defect recognition technology, and is a method and system for classifying morphological defects in silicon carbide epitaxial layers based on reinforcement learning. Background Technology
[0002] Silicon carbide (SiC), as a third-generation wide-bandgap semiconductor material, has irreplaceable application prospects in high-voltage, high-frequency, and high-temperature power devices. The quality of the epitaxial layer directly determines the breakdown voltage and reliability of the device. During the critical window period between the end of epitaxial growth and the beginning of chemical mechanical polishing, the SiC epitaxial layer needs to undergo a rapid cooling process from its growth temperature of approximately 1600°C to room temperature. During this process, due to the mismatch in thermal expansion coefficients between the substrate and the epitaxial layer, the doping concentration gradient, and the inherent anisotropy of the 4H-SiC crystal, significant non-uniform thermal stress relaxation phenomena are induced, which in turn drive the slippage, extension, or morphological evolution of microscopic defects such as dislocations and stacking faults. Existing technologies mostly employ static optical morphology detection at room temperature, i.e., acquiring images after the defects have fully shaped and classifying them based on convolutional neural networks. However, this type of method has two fundamental limitations. First, the surface of SiC epitaxial layers naturally exhibits step flow textures and step clustering structures induced by 4° off-cut growth. These normal growth features, under specific lighting conditions, produce scattering artifacts highly similar to cracks or scratches. Static classification models are prone to misclassifying these as defects, leading to severe false positives. Second, SiC material has a high refractive index of 2.6 and strong birefringence, resulting in surface morphology distortions induced by numerous deep dislocations with a thickness of only a few micrometers. Surface contamination or oxide residues can also produce similar optical signals. Two-dimensional images alone cannot distinguish the true depth and physical nature of defects. Existing methods struggle to determine the physical nature of morphological anomalies without relying on expensive equipment such as laser scanning. Furthermore, during the cooling thermal stress relaxation process, the morphology of the defect is not static but undergoes transient evolution with the temperature gradient. The same defect may exhibit completely different optical characteristics at different cooling stages. Traditional detection methods only collect static images at a single time point, losing the dynamic information of defect evolution, resulting in a lack of predictive ability to determine whether the defect will expand into a failure source in subsequent processes (such as ion implantation and high-temperature annealing). Summary of the Invention
[0003] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.
[0004] The technical problem to be solved by the present invention is that it is difficult to accurately distinguish the normal growth artifacts caused by the 4° off-axis step flow under dynamic thermal stress environment in the prior art. The present invention proposes a method and system for classifying the morphological defects of silicon carbide epitaxial layers based on reinforcement learning.
[0005] To achieve the above objectives, the technical solution of the silicon carbide epitaxial layer morphology defect classification method based on reinforcement learning of the present invention includes the following steps:
[0006] Acquire historical morphological evolution data of silicon carbide epitaxial testing equipment, lattice dislocation topology matrix parameters of epitaxial layer surface, and multi-view morphological response tensor triggered by cooling thermal stress coupling;
[0007] Based on the historical morphological evolution data, the morphological feature mapping stability parameter for defect identification is obtained, and the semantic convergence limit for target classification is calculated using the morphological feature mapping stability parameter.
[0008] The primary light intensity distribution vector of the epitaxial layer to be tested is extracted, and a defect discrimination evolution model based on reinforcement learning is constructed. The primary light intensity distribution vector, the lattice dislocation topology matrix parameters, and the multi-view morphology response tensor are imported into the defect discrimination evolution model to iteratively generate multiple sets of spatial dimension feature characterization values.
[0009] The confidence level of the multiple sets of spatial dimension feature representation values is fitted and filtered by the target classification semantic convergence boundary to obtain the target execution action command, and drive the silicon carbide morphology detection module to perform attitude registration mapping on the physical space vector of the probe beam, and output the epitaxial layer morphology defect classification result.
[0010] Preferably, the physical space vector of the probe beam is characterized by a multi-level spatial domain analysis subsystem: each level of the spatial domain analysis subsystem is composed of four sets of characteristic response operators, wherein the first response operator and the third response operator are distributed along the 4° off-axis projection axis of the SiC crystal plane, and the second response operator and the fourth response operator are distributed parallel to the SiC step growth texture.
[0011] The first response operator, the second response operator, and the fourth response operator are detection operators with a fixed birefringence compensation rate; the third response operator is a dynamic gain adjustment operator controlled by a reinforcement learning agent and capable of performing spatial angle deflection actions.
[0012] The characteristic light field flows through the multi-level spatial domain analysis subsystem, generating the first morphological feature point, the second morphological feature point, the third morphological feature point, and the fourth morphological feature point at the four sets of response operators in sequence;
[0013] Among them, the first morphological feature point, the second morphological feature point, and the third morphological feature point are used to perform feature fusion and background suppression processing on stray signals containing cooling gradient noise; the fourth morphological feature point is used to perform normalized projection processing on the fused signal for the final classification decision.
[0014] Preferably, the lattice dislocation topology matrix parameters include: step feature width component. Initial feature sampling step size and terminal reconstruction projection scale ;
[0015] The multi-view topography response tensor includes: a dynamic feature displacement mapping set element. ;
[0016] The extraction method for the dynamic feature displacement mapping set element is as follows:
[0017] S11: Identify the physical location coordinates of the dislocation edge captured by the second response operator on the topographic projection;
[0018] S12: Fine-tune the probe axis of the third response operator to make it collinear with the current SiC step drift vector direction affected by thermal stress;
[0019] S13: After initial collinear alignment is completed, based on the reinforcement learning exploration mechanism, the third response operator is controlled to perform a feature detection angle rotation, so that the outgoing light wave from the second response operator acts on the third response operator with different coherent interference deflection angles. The deflection angle adjustment range is... ;
[0020] S14: Real-time recording of rotation angle Dynamically offset multidimensional feature mapping trajectory points generate transient topographic projection clusters;
[0021] S15: Calculate the coherence path difference between the initial identification coordinates and each mapping point within the projection cluster in the signal space, and construct a dynamic feature displacement mapping set element. ,in The total number of nodes is determined by the sampling accuracy; the dynamic feature displacement mapping set element is used to calculate the feature refinement depth of the double sampling accuracy.
[0022] Preferably, the multi-view topography response tensor further includes: a set of coupled evolutionary displacement vectors. ;
[0023] The acquisition of the coupled evolution displacement vector set includes:
[0024] S16: Based on the SiC epitaxial growth kinetics model, simulate the geometric distortion when the feature elements in the transient topography projection cluster migrate to the fourth response operator's detection window due to thermal relaxation.
[0025] S17: Synchronously acquire the topological consistency measurement values between each feature element and the generated distortion position to form the coupled evolution displacement vector set. ;
[0026] The coupled evolution displacement vector set is used to calculate the triple sampling feature contribution depth.
[0027] Preferably, the morphological feature mapping stability parameter is calculated as follows:
[0028] S21: Extract historical training trajectory data from the historical inspection database, covering SiC batch defect rate, strain field residual coefficient, and classification accuracy gain. The historical training trajectory data includes normalized defect feature response intensity. Semantic information gain of normalized classification features ,in For sampling index;
[0029] S22: Calculate the correlation coefficient of defect feature evolution based on the historical training trajectory data, and output the morphological feature mapping stability parameter. :
[0030] ;
[0031] in, For training rounds, For single-sampling index; The intensity of the defect characteristic response during detection; The semantic information gain of the classification features at this moment; , This is the corresponding average value;
[0032] S23: Mapping morphological features to stability parameters Import the convergence bound generator to obtain the target classification semantic convergence bound for classification decisions. The specific formula is as follows:
[0033] ;
[0034] in, The basic classification and discrimination entropy is preset based on the thickness of the SiC epitaxial layer. The natural exponential function is used; the target classification semantic convergence bound is used to perform confidence fitting and screening of multiple sets of spatial dimension feature representation values.
[0035] Preferably, the defect discrimination evolution model incorporates a strategy for extrapolating the flow of silicon carbide epitaxial morphology features:
[0036] S31: Construct a multi-order topographic feature scattering evolution function and define the feature transfer coefficients. for:
[0037] ;
[0038] in, The optical interference efficiency coefficient; This represents the geometric bias guided by a 4° off-axis angle of the SiC wafer; The angle between the defect profile and the probe beam; The normalized compensation factor for birefringence of 4H hexagonal crystal; the characteristic transfer coefficient Used to recursively calculate the equivalent feature transmission distance at each response operator;
[0039] S32: Utilizing the step feature width component and initial equivalent characteristic transmission distance Calculate the contribution depth of the primary feature to obtain the primary recognition features. The calculation formula is:
[0040] ;
[0041] in, The equivalent characteristic transmission distance corresponding to the primary light intensity distribution vector is... This is the preset normalization reference factor;
[0042] S33: Contribute the first-level recognition features to the depth. Import the discriminant model and calculate the feature transfer coefficient of the operator based on the state parameters at the second response operator. And predict the equivalent feature transfer distance when the feature flow is transferred to the second response operator. Then, the depth of the first-level classification information at this dimension is calculated. The calculation formula is:
[0043] ;
[0044] in, This is the initial feature sampling step size;
[0045] S34: Perform cascaded propagation of the identification results, and calculate the feature propagation coefficient of the third response operator based on the state parameters at the third response operator. The equivalent feature transmission distance when the defect evolution flow is projected from the second observation dimension to the third dimension is obtained. And based on the dynamic characteristic displacement mapping set element Generate a feature contribution depth set with double sampling precision
[0046] ,in: ;
[0047] in, , which is an element in the set of elements of the dynamic feature displacement mapping.
[0048] Preferably, the defect discrimination evolution model further includes:
[0049] S35: Adjustable deflection angle at the third response operator extraction point The introduced coupling interference suppression amount is used to calculate the characteristic transfer coefficient of the operator based on the state parameters at the fourth response operator. The equivalent feature transfer distance when the feature flow is projected onto the fourth response operator is obtained. And determine the depth set of triple sampling feature contributions under the divide-and-conquer model from each angle. ,in: ;
[0050] in, These are the elements in the set of coupled evolution displacement vectors;
[0051] S36: Converged Terminal Reconstructs Projection Scale The defect classification logic flow is obtained by mapping the exit mapping feature contribution depth set from the detector's final sensing end to the discrimination probability space. ,in: .
[0052] Preferably, the extraction process of the multiple sets of spatial dimension feature representation values is as follows:
[0053] S41: Aggregates all feature mapping values generated within the recognition cycle, including the contribution depth of the primary feature. First-level feature contribution depth Double sampling feature contribution depth set Triple sampling feature contribution depth set and the depth set of export mapping feature contributions ;
[0054] S42: Establish a criterion for calculating the spatial logical total representation depth under the global reward function of reinforcement learning, and obtain the set of spatial dimension feature representation values. ,in: ;
[0055] in, , , , , All of these are feature contribution depths calculated from the aforementioned steps.
[0056] Preferably, the semantic convergence limit of the target classification is used. The specific operations for filtering multiple sets of spatial dimension feature representation values include:
[0057] Set the compensation amount for the characteristic characterization deviation caused by the SiC step-bundling effect exfoliation. The feature characterizes the deviation compensation amount It is a preset positive number used to compensate for the feature representation deviation caused by step clustering artifacts;
[0058] Calculate the semantic convergence bound of target classification Compensation amount for deviation from the aforementioned feature representation The sum of values is used to construct a safe discrimination threshold for classification and recognition. ;
[0059] Set of spatial dimensional feature representation values Each component within With security discrimination threshold Perform deviation calculation to calculate the feature evolution fit difference. The fitted residuals are arranged in ascending order, and the top 5%-10% of action sequences with the best characteristics are identified as the target classification scheme set.
[0060] Analyze the deviation angle of the corresponding position of the third response operator in the target classification scheme set. Matching the optimal action decision in reinforcement learning;
[0061] The system senses the current physical field of view of the photosensitive component and sends an angle adjustment command to the drive mechanism to dynamically adjust the spatial observation angle of the third response operator.
[0062] In addition, the silicon carbide epitaxial layer morphology defect classification system based on reinforcement learning of this invention includes the following modules:
[0063] Microscopic multidimensional scanning analysis unit, evolutionary boundary decision unit, autoregressive classification control engine, and intelligent action evaluator;
[0064] The microscopic multidimensional scanning analysis unit is used to acquire the crystallographic projection of the SiC substrate in the post-growth stage and its stress-triggered transient scattering pattern in real time, and to construct the lattice dislocation topology matrix parameters and the multi-view topology response tensor based on the step flow dynamics parameters.
[0065] The evolution boundary decision unit adjusts the classification convergence logic based on historical detection feedback, calculates the target classification semantic convergence boundary through the morphological feature mapping stability parameter, and establishes a feature discrimination stopping boundary aimed at eliminating step clustering false positives.
[0066] The autoregressive classification control engine includes a multi-order feature flow simulator and an intelligent action evaluator. The multi-order feature flow simulator has a built-in multi-order morphology feature scattering evolution function, which is used to recursively calculate the equivalent feature transmission distance at each response operator.
[0067] The intelligent action evaluator is used to output the target execution action command based on the filtering results of the set of spatial dimension feature representation values and the target classification semantic convergence boundary, drive the silicon carbide morphology detection module to perform attitude registration mapping on the physical space vector of the probe beam, and complete the adaptive and accurate classification execution of SiC epitaxial morphology anomalies.
[0068] Compared with the prior art, the technical effects of the present invention are as follows:
[0069] 1. This invention enables the differentiation between step-bundling artifacts and real defects, solving the problem of excessively high false positive rates in existing static inspection. Because the naturally formed step flow textures during SiC epitaxial growth at a 4° off-axis angle produce scattering signals highly similar to cracks and scratches under specific optical conditions, traditional classification models cannot distinguish between normal growth features and potential failure defects. This leads to a large number of qualified epitaxial wafers being misclassified as defective, severely reducing production yield and increasing unnecessary process costs. This invention constructs a multi-level spatial domain analysis subsystem and introduces a reinforcement learning agent, enabling the third response operator to dynamically perform spatial angle deflection based on the currently detected morphological features. It actively changes the coherent interference deflection angle to obtain multi-view morphological response tensors from different crystal orientations, thereby decoupling the artifact signals generated by step-bundling from the real defect signals induced by dislocations in physical space. This achieves synergistic optimization of background suppression and feature enhancement, improving the confidence and reliability of the classification results.
[0070] 2. SiC material has high refractive index and strong birefringence, making it difficult to distinguish the morphological features of surface contamination and deep dislocations from a single perspective. However, this invention expands the feature contribution depth from a single dimension to a multi-level recursive spatial dimension feature characterization value set, enabling the system to distinguish whether the defect is a physical scratch that only exists on the surface or a deep dislocation type defect that has the risk of extending into the active region of the device. This provides a predictive basis for whether the defect will expand into a fatal failure source in subsequent processes.
[0071] In summary, this solution, without adding high-cost hardware such as laser scanning, achieves both a reduction in false positive rate and accurate identification of the physical nature of defects by simply optimizing the optical detection posture and fusing analysis of the contribution depth of multi-scale features. This effectively improves the process adaptability of silicon carbide epitaxial layer detection and the long-term reliability of the final device. Attached Figure Description
[0072] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:
[0073] Figure 1 This is a flowchart illustrating the reinforcement learning-based silicon carbide epitaxial layer morphology defect classification method of the present invention.
[0074] Figure 2 This is a schematic diagram of the structure of the silicon carbide epitaxial layer morphology defect classification system based on reinforcement learning of the present invention. Detailed Implementation
[0075] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0076] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0077] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.
[0078] Example 1:
[0079] like Figure 1 As shown in the figure, the reinforcement learning-based silicon carbide epitaxial layer morphology defect classification method of this invention is as follows: Figure 1 As shown, the specific steps include the following:
[0080] Acquire historical morphological evolution data of silicon carbide epitaxial testing equipment, lattice dislocation topology matrix parameters of epitaxial layer surface, and multi-view morphological response tensor triggered by cooling thermal stress coupling;
[0081] The physical space vector of the probe beam is characterized by a multi-level spatial analysis subsystem: each level of spatial analysis subsystem is composed of four sets of characteristic response operators, wherein the first response operator and the third response operator are distributed along the 4° off-axis projection axis of the SiC(0001) crystal plane, and the second response operator and the fourth response operator are distributed parallel to the SiC step growth texture.
[0082] The first response operator, the second response operator, and the fourth response operator are detection operators with a fixed birefringence compensation rate; the third response operator is a dynamic gain adjustment operator controlled by a reinforcement learning agent and capable of performing spatial angle deflection actions.
[0083] The characteristic light field flows through the multi-level spatial domain analysis subsystem, generating the first morphological feature point, the second morphological feature point, the third morphological feature point, and the fourth morphological feature point at the four sets of response operators in sequence;
[0084] Among them, the first morphological feature point, the second morphological feature point, and the third morphological feature point are used to perform feature fusion and background suppression processing on stray signals containing cooling gradient noise; the fourth morphological feature point is used to perform normalized projection processing on the fused signal for the final classification decision.
[0085] The lattice dislocation topology matrix parameters include: step feature width component. Initial feature sampling step size and terminal reconstruction projection scale ;
[0086] The multi-view topography response tensor includes: a dynamic feature displacement mapping set element. ;
[0087] The extraction method for the dynamic feature displacement mapping set element is as follows:
[0088] S11: Identify the physical location coordinates of the dislocation edge captured by the second response operator on the topographic projection;
[0089] S12: Fine-tune the probe axis of the third response operator to make it collinear with the current SiC step drift vector direction affected by thermal stress;
[0090] S13: After initial collinear alignment is completed, based on the reinforcement learning exploration mechanism, the third response operator is controlled to perform a feature detection angle rotation, so that the outgoing light wave from the second response operator acts on the third response operator with different coherent interference deflection angles. The deflection angle adjustment range is... ;
[0091] S14: Real-time recording of rotation angle Dynamically offset multidimensional feature mapping trajectory points generate transient topographic projection clusters;
[0092] S15: Calculate the coherence path difference between the initial identification coordinates and each mapping point within the projection cluster in the signal space, and construct a dynamic feature displacement mapping set element. ,in The total number of nodes is determined by the sampling accuracy; the dynamic feature displacement mapping set element is used to calculate the feature refinement depth of the double sampling accuracy.
[0093] The multi-view topography response tensor also includes: a set of coupled evolutionary displacement vectors. ;
[0094] The acquisition of the coupled evolution displacement vector set includes:
[0095] S16: Based on the SiC epitaxial growth kinetics model, simulate the geometric distortion when the feature elements in the transient topography projection cluster migrate to the fourth response operator's detection window due to thermal relaxation.
[0096] S17: Synchronously acquire the topological consistency measurement values between each feature element and the generated distortion position to form the coupled evolution displacement vector set. ;
[0097] The coupled evolution displacement vector set is used to calculate the triple sampling feature contribution depth.
[0098] Based on the historical morphological evolution data, the morphological feature mapping stability parameter for defect identification is obtained, and the semantic convergence limit for target classification is calculated using the morphological feature mapping stability parameter.
[0099] The morphological feature mapping stability parameter is calculated as follows:
[0100] S21: Extract historical training trajectory data from the historical inspection database, covering SiC batch defect rate, strain field residual coefficient, and classification accuracy gain. The historical training trajectory data includes normalized defect feature response intensity. Semantic information gain of normalized classification features ,in For sampling index;
[0101] S22: Calculate the correlation coefficient of defect feature evolution based on the historical training trajectory data, and output the morphological feature mapping stability parameter. :
[0102] ;
[0103] in, For training rounds, For single-sampling index; The intensity of the defect characteristic response during detection; The semantic information gain of the classification features at this moment; , This represents the corresponding average value;
[0104] S23: Mapping morphological features to stability parameters Import the convergence bound generator to obtain the target classification semantic convergence bound for classification decisions. The specific formula is as follows:
[0105] ;
[0106] in, The basic classification and discrimination entropy is preset based on the thickness of the SiC epitaxial layer. The natural exponential function is used; the target classification semantic convergence bound is used to perform confidence fitting and screening of multiple sets of spatial dimension feature representation values.
[0107] The primary light intensity distribution vector of the epitaxial layer to be tested is extracted, and a defect discrimination evolution model based on reinforcement learning is constructed. The primary light intensity distribution vector, the lattice dislocation topology matrix parameters, and the multi-view morphology response tensor are imported into the defect discrimination evolution model to iteratively generate multiple sets of spatial dimension feature characterization values.
[0108] The defect discrimination evolution model has a built-in strategy for inferring the flow of silicon carbide epitaxial morphology features:
[0109] S31: Construct a multi-order topographic feature scattering evolution function and define the feature transfer coefficients. for:
[0110] ;
[0111] in, The optical interference efficiency coefficient; This represents the geometric bias guided by a 4° off-axis angle of the SiC wafer; The angle between the defect profile and the probe beam; The normalized compensation factor for birefringence of 4H hexagonal crystal; the characteristic transfer coefficient Used to recursively calculate the equivalent feature transmission distance at each response operator;
[0112] S32: Utilizing the step feature width component and initial equivalent characteristic transmission distance Calculate the contribution depth of the primary feature to obtain the primary recognition features. The calculation formula is:
[0113] ;
[0114] in, The equivalent characteristic transmission distance corresponding to the primary light intensity distribution vector is... This is the preset normalization reference factor;
[0115] S33: Contribute the first-level recognition features to the depth. Import the discriminant model and calculate the feature transfer coefficient of the operator based on the state parameters at the second response operator. And predict the equivalent feature transfer distance when the feature flow is transferred to the second response operator. Then, the depth of the first-level classification information at this dimension is calculated. The calculation formula is:
[0116] ;
[0117] in, This is the initial feature sampling step size;
[0118] S34: Perform cascaded propagation of the identification results, and calculate the feature propagation coefficient of the third response operator based on the state parameters at the third response operator. The equivalent feature transmission distance when the defect evolution flow is projected from the second observation dimension to the third dimension is obtained. And based on the dynamic characteristic displacement mapping set element Generate a feature contribution depth set with double sampling precision
[0119] ,in: ;
[0120] in, , which is an element in the set of elements of the dynamic feature displacement mapping.
[0121] The defect discrimination evolution model also includes:
[0122] S35: Adjustable deflection angle at the third response operator extraction point The introduced coupling interference suppression amount is based on the state parameters at the fourth response operator (including the adjustable deflection angle). Calculate the characteristic transfer coefficients of this operator. The equivalent feature transfer distance when the feature flow is projected onto the fourth response operator is obtained. And determine the depth set of triple sampling feature contributions under the divide-and-conquer model from each angle. ,in: ;
[0123] in, These are the elements in the set of coupled evolution displacement vectors;
[0124] S36: Converged Terminal Reconstructs Projection Scale The defect classification logic flow is obtained by mapping the exit mapping feature contribution depth set from the detector's final sensing end to the discrimination probability space. ,in: .
[0125] The confidence level of the multiple sets of spatial dimension feature representation values is fitted and filtered by the target classification semantic convergence boundary to obtain the target execution action command, and drive the silicon carbide morphology detection module to perform attitude registration mapping on the physical space vector of the probe beam, and output the epitaxial layer morphology defect classification result.
[0126] The extraction process of the multiple sets of spatial dimension feature representation values is as follows:
[0127] S41: Aggregates all feature mapping values generated within the recognition cycle, including the contribution depth of the primary feature. First-level feature contribution depth Double sampling feature contribution depth set Triple sampling feature contribution depth set and the depth set of export mapping feature contributions ;
[0128] S42: Establish a criterion for calculating the spatial logical total representation depth under the global reward function of reinforcement learning, and obtain the set of spatial dimension feature representation values. ,in: ;
[0129] in, , , , , All of these are feature contribution depths calculated from the aforementioned steps.
[0130] Convergence bounds based on target classification semantics The specific operations for filtering multiple sets of spatial dimension feature representation values include:
[0131] Set the compensation amount for the characteristic characterization deviation caused by the SiC step-bundling effect exfoliation. The feature characterizes the deviation compensation amount It is a preset positive number used to compensate for the feature representation deviation caused by step clustering artifacts;
[0132] Calculate the semantic convergence bound of target classification Compensation amount for deviation from the aforementioned feature representation The sum of values is used to construct a safe discrimination threshold for classification and recognition. ;
[0133] Set of spatial dimensional feature representation values Each component within With security discrimination threshold Perform deviation calculation to calculate the feature evolution fit difference. The fitted residuals are arranged in ascending order, and the top 5%-10% of action sequences with the best characteristics are identified as the target classification scheme set.
[0134] Analyze the deviation angle of the corresponding position of the third response operator in the target classification scheme set. Matching the optimal action decision in reinforcement learning;
[0135] The system senses the current physical field of view of the photosensitive component and sends an angle adjustment command to the drive mechanism to dynamically adjust the spatial observation angle of the third response operator. This is done to reduce background artifact interference and achieve physical morphology classification of defects.
[0136] For example, in one specific implementation, the step of outputting the classification results of epitaxial layer morphology defects specifically includes:
[0137] After the autoregressive classification control engine generates the target execution action command and completes the physical space vector attitude registration mapping of the probe beam, the system collects the normalized projection signal at the fourth response operator after attitude optimization in real time, inputs the signal into the preset defect type decoder, and the decoder maps the defect category label corresponding to the current detection area based on the fitting and filtering results of the feature contribution depth set and the target classification semantic convergence boundary.
[0138] The defect categories include at least step-bundle defects, triangular defects, and dislocation defects;
[0139] Subsequently, the system outputs the classification results in the form of structured data, which includes defect type identifier, two-dimensional position coordinates of the defect in the wafer coordinate system, and classification confidence score, and sends it to the manufacturing execution system simultaneously to trigger defective product diversion.
[0140] Meanwhile, the system also stores the classification results of the current frame and the corresponding pose adjustment parameters into the historical detection database for rolling updates of morphological feature mapping stability parameters, thereby achieving self-evolution of the classification boundary.
[0141] Example 2:
[0142] like Figure 2 As shown in the figure, the reinforcement learning-based silicon carbide epitaxial layer morphology defect classification system of this invention is as follows: Figure 2 As shown, it includes the following modules:
[0143] Microscopic multidimensional scanning analysis unit, evolutionary boundary decision unit, autoregressive classification control engine, and intelligent action evaluator;
[0144] The microscopic multidimensional scanning analysis unit is used to acquire the crystallographic projection of the SiC substrate in the post-growth stage and its stress-triggered transient scattering pattern in real time, and to construct the lattice dislocation topology matrix parameters and the multi-view topology response tensor based on the step flow dynamics parameters.
[0145] The evolution boundary decision unit adjusts the classification convergence logic based on historical detection feedback, calculates the target classification semantic convergence boundary through the morphological feature mapping stability parameter, and establishes a feature discrimination stopping boundary aimed at eliminating step clustering false positives.
[0146] The autoregressive classification control engine includes a multi-order feature flow simulator and an intelligent action evaluator. The multi-order feature flow simulator has a built-in multi-order morphology feature scattering evolution function, which is used to recursively calculate the equivalent feature transmission distance at each response operator.
[0147] The intelligent action evaluator is used to output the target execution action command based on the filtering results of the set of spatial dimension feature representation values and the target classification semantic convergence boundary, drive the silicon carbide morphology detection module to perform attitude registration mapping on the physical space vector of the probe beam, and complete the adaptive and accurate classification execution of SiC epitaxial morphology anomalies.
[0148] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired network and / or wireless network. A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives (SSDs).
[0149] In the several embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only one method, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.
[0150] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0151] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
Claims
1. A method for classifying morphological defects in silicon carbide epitaxial layers based on reinforcement learning, characterized in that, The method includes: Acquire historical morphological evolution data of silicon carbide epitaxial testing equipment, lattice dislocation topology matrix parameters of epitaxial layer surface, and multi-view morphological response tensor triggered by cooling thermal stress coupling; Based on the historical morphological evolution data, the morphological feature mapping stability parameter for defect identification is obtained, and the semantic convergence limit for target classification is calculated using the morphological feature mapping stability parameter. The primary light intensity distribution vector of the epitaxial layer to be tested is extracted, and a defect discrimination evolution model based on reinforcement learning is constructed. The primary light intensity distribution vector, the lattice dislocation topology matrix parameters, and the multi-view morphology response tensor are imported into the defect discrimination evolution model to iteratively generate multiple sets of spatial dimension feature characterization values. The confidence level of the multiple sets of spatial dimension feature representation values is fitted and filtered by the target classification semantic convergence boundary to obtain the target execution action command, and drive the silicon carbide morphology detection module to perform attitude registration mapping on the physical space vector of the probe beam, and output the epitaxial layer morphology defect classification result.
2. The method for classifying morphological defects in silicon carbide epitaxial layers based on reinforcement learning according to claim 1, characterized in that, The physical space vector of the probe beam is characterized by a multi-level spatial analysis subsystem: each level of spatial analysis subsystem is composed of four sets of characteristic response operators, wherein the first response operator and the third response operator are distributed along the 4° off-axis projection axis of the SiC crystal plane, and the second response operator and the fourth response operator are distributed parallel to the SiC step growth texture. The first response operator, the second response operator, and the fourth response operator are detection operators with a fixed birefringence compensation rate; the third response operator is a dynamic gain adjustment operator controlled by a reinforcement learning agent and capable of performing spatial angle deflection actions. The characteristic light field flows through the multi-level spatial domain analysis subsystem, generating the first morphological feature point, the second morphological feature point, the third morphological feature point, and the fourth morphological feature point at the four sets of response operators in sequence; Among them, the first morphological feature point, the second morphological feature point, and the third morphological feature point are used to perform feature fusion and background suppression processing on stray signals containing cooling gradient noise; the fourth morphological feature point is used to perform normalized projection processing on the fused signal for the final classification decision.
3. The method for classifying morphological defects in silicon carbide epitaxial layers based on reinforcement learning according to claim 2, characterized in that, The lattice dislocation topology matrix parameters include: step feature width component. Initial feature sampling step size and terminal reconstruction projection scale ; The multi-view topography response tensor includes: a dynamic feature displacement mapping set element. ; The extraction method for the dynamic feature displacement mapping set element is as follows: S11: Identify the physical location coordinates of the dislocation edge captured by the second response operator on the topographic projection; S12: Fine-tune the probe axis of the third response operator to make it collinear with the current SiC step drift vector direction affected by thermal stress; S13: After initial collinear alignment is completed, based on the reinforcement learning exploration mechanism, the third response operator is controlled to perform a feature detection angle rotation, so that the outgoing light wave from the second response operator acts on the third response operator with different coherent interference deflection angles. The deflection angle adjustment range is... ; S14: Real-time recording of rotation angle Dynamically offset multidimensional feature mapping trajectory points generate transient topographic projection clusters; S15: Calculate the coherence path difference between the initial identification coordinates and each mapping point within the projection cluster in the signal space, and construct a dynamic feature displacement mapping set element. ,in The total number of nodes is determined by the sampling accuracy; the dynamic feature displacement mapping set element is used to calculate the feature refinement depth of the double sampling accuracy.
4. The method for classifying morphological defects in silicon carbide epitaxial layers based on reinforcement learning according to claim 3, characterized in that, The multi-view topography response tensor also includes: a set of coupled evolutionary displacement vectors. ; The acquisition of the coupled evolution displacement vector set includes: S16: Based on the SiC epitaxial growth kinetics model, simulate the geometric distortion when the feature elements in the transient topography projection cluster migrate to the detection window of the fourth response operator due to thermal relaxation. S17: Synchronously acquire the topological consistency measurement values between each feature element and the generated distortion position to form the coupled evolution displacement vector set. ; The coupled evolution displacement vector set is used to calculate the triple sampling feature contribution depth.
5. The method for classifying morphological defects in silicon carbide epitaxial layers based on reinforcement learning according to claim 4, characterized in that, The morphological feature mapping stability parameter is calculated as follows: S21: Extract historical training trajectory data from the historical inspection database, covering SiC batch defect rate, strain field residual coefficient, and classification accuracy gain. The historical training trajectory data includes normalized defect feature response intensity. Semantic information gain of normalized classification features ,in For sampling index; S22: Calculate the correlation coefficient of defect feature evolution based on the historical training trajectory data, and output the morphological feature mapping stability parameter. : ； in, For training rounds, For single-sampling index; The intensity of the defect characteristic response during detection; The semantic information gain of the classification features at this moment; , This is the corresponding average value; S23: Mapping morphological features to stability parameters Import the convergence bound generator to obtain the target classification semantic convergence bound for classification decisions. The specific formula is: ； in, The basic classification and discrimination entropy is preset based on the thickness of the SiC epitaxial layer. The natural exponential function is used; the target classification semantic convergence bound is used to perform confidence fitting and screening of multiple sets of spatial dimension feature representation values.
6. The method for classifying morphological defects in silicon carbide epitaxial layers based on reinforcement learning according to claim 5, characterized in that, The defect discrimination evolution model has a built-in strategy for inferring the flow of silicon carbide epitaxial morphology features: S31: Construct a multi-order topographic feature scattering evolution function and define the feature transfer coefficients. for: ； in, The optical interference efficiency coefficient; This represents the geometric bias guided by a 4° off-axis angle of the SiC wafer; The angle between the defect profile and the probe beam; The normalized compensation factor for birefringence of 4H hexagonal crystal; the characteristic transfer coefficient Used to recursively calculate the equivalent feature transmission distance at each response operator; S32: Utilizing the step feature width component and initial equivalent characteristic transmission distance Calculate the contribution depth of the primary feature to obtain the primary recognition features. The calculation formula is: ； in, The equivalent characteristic transmission distance corresponding to the primary light intensity distribution vector is... This is the preset normalization reference factor; S33: Contribute the first-level recognition features to the depth. Import the discriminant model and calculate the feature transfer coefficient of the operator based on the state parameters at the second response operator. And predict the equivalent feature transfer distance when the feature flow is transferred to the second response operator. Then, the depth of the first-level classification information at this dimension is calculated. The calculation formula is: ； in, This is the initial feature sampling step size; S34: Perform cascaded propagation of the identification results, and calculate the feature propagation coefficient of the third response operator based on the state parameters at the third response operator. The equivalent feature transmission distance when the defect evolution flow is projected from the second observation dimension to the third dimension is obtained. And based on the dynamic characteristic displacement mapping set element Generate a feature contribution depth set with double sampling precision ,in: ; in, , which is an element in the set of elements of the dynamic feature displacement mapping.
7. The method for classifying morphological defects in silicon carbide epitaxial layers based on reinforcement learning according to claim 6, characterized in that, The defect discrimination evolution model also includes: S35: Adjustable deflection angle at the third response operator extraction point The introduced coupling interference suppression amount is used to calculate the characteristic transfer coefficient of the operator based on the state parameters at the fourth response operator. The equivalent feature transfer distance when the feature flow is projected onto the fourth response operator is obtained. And determine the depth set of triple sampling feature contributions under the divide-and-conquer model from each angle. ,in: ; in, These are the elements in the set of coupled evolution displacement vectors; S36: Converged Terminal Reconstructs Projection Scale The defect classification logic flow is obtained by mapping the exit mapping feature contribution depth set from the detector's final sensing end to the discrimination probability space. ,in: .
8. The method for classifying morphological defects in silicon carbide epitaxial layers based on reinforcement learning according to claim 7, characterized in that, The extraction process of the multiple sets of spatial dimension feature representation values is as follows: S41: Aggregates all feature mapping values generated within the recognition cycle, including the contribution depth of the primary feature. First-level feature contribution depth Double sampling feature contribution depth set Triple sampling feature contribution depth set and the deep set of export mapping feature contributions ; S42: Establish a criterion for calculating the spatial logical total representation depth under the global reward function of reinforcement learning, and obtain the set of spatial dimension feature representation values. ,in: ; in, , , , , All of these are feature contribution depths calculated from the aforementioned steps.
9. The method for classifying morphological defects in silicon carbide epitaxial layers based on reinforcement learning according to claim 8, characterized in that, Convergence bounds based on target classification semantics The specific operations for filtering multiple sets of spatial dimension feature representation values include: Set the compensation amount for the characteristic characterization deviation caused by the SiC step-bundling effect exfoliation. The feature characterization deviation compensation amount It is a preset positive number used to compensate for the feature representation deviation caused by step clustering artifacts; Calculate the semantic convergence bound of target classification Compensation amount for deviation from the aforementioned feature representation The sum of values is used to construct a safe discrimination threshold for classification and recognition. ; Set of spatial dimensional feature representation values Each component within With security discrimination threshold Perform deviation calculation to calculate the feature evolution fit difference. The fitted residuals are arranged in ascending order, and the top 5%-10% of action sequences with the best characteristics are identified as the target classification scheme set. Analyze the deviation angle of the corresponding position of the third response operator in the target classification scheme set. Matching the optimal action decision in reinforcement learning; The system senses the current physical field of view of the photosensitive component and sends an angle adjustment command to the drive mechanism to dynamically adjust the spatial observation angle of the third response operator.
10. A reinforcement learning-based silicon carbide epitaxial layer morphology defect classification system, used to implement the reinforcement learning-based silicon carbide epitaxial layer morphology defect classification method as described in any one of claims 6-9, characterized in that, The system includes: Microscopic multidimensional scanning analysis unit, evolutionary boundary decision unit, autoregressive classification control engine, and intelligent action evaluator; The microscopic multidimensional scanning analysis unit is used to acquire the crystallographic projection of the SiC substrate in the post-growth stage and its stress-triggered transient scattering pattern in real time, and to construct the lattice dislocation topology matrix parameters and the multi-view topology response tensor based on the step flow dynamics parameters. The evolution boundary decision unit adjusts the classification convergence logic based on historical detection feedback, calculates the target classification semantic convergence boundary through the morphological feature mapping stability parameter, and establishes a feature discrimination stopping boundary aimed at eliminating step clustering false positives. The autoregressive classification control engine includes a multi-order feature flow simulator and an intelligent action evaluator. The multi-order feature flow simulator has a built-in multi-order morphology feature scattering evolution function, which is used to recursively calculate the equivalent feature transmission distance at each response operator. The intelligent action evaluator is used to output the target execution action command based on the filtering results of the set of spatial dimension feature representation values and the target classification semantic convergence boundary, drive the silicon carbide morphology detection module to perform attitude registration mapping on the physical space vector of the probe beam, and complete the adaptive and accurate classification execution of SiC epitaxial morphology anomalies.