Semiconductor defect detection system based on image recognition
By constructing multi-time response sequences synchronized with thermal history and using time-series deep learning modeling, the problem of ambiguous defect detection features under unsteady conditions in traditional methods is solved, achieving stable characterization and accurate identification of semiconductor defects and improving the reliability of the detection system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANDONG JIANZHU UNIV
- Filing Date
- 2026-04-22
- Publication Date
- 2026-05-29
Smart Images
Figure CN122115423A_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to the field of semiconductor manufacturing technology, and in particular to a semiconductor defect detection method and system based on image recognition. Background Technology
[0002] In semiconductor wafer manufacturing, rapid and accurate online detection of defects is crucial for achieving high yield and reliability. The fractional Fourier transform, due to its ability to provide a unified time-frequency domain representation, has been applied to feature extraction and characterization of defect responses. However, in some critical processes, especially under conditions where online detection is performed immediately after rapid thermal processing, the wafer is in a significant unsteady thermal history phase. Its internal thermal stress distribution, carrier concentration gradient, and surface thin film optical properties continue to evolve rapidly over time. In this unsteady condition, the optical or electrical responses induced by defects exhibit transient processes that continuously drift over time.
[0003] Traditional fractional Fourier transform-based methods implicitly rely on a quasi-steady-state assumption when constructing a unified time-frequency representation, assuming that the statistical characteristics of the target signal's time-frequency structure are stable within the detection time window. This premise no longer holds when the wafer is under severe thermal fluctuations. The energy distribution of the defect response in the time-frequency domain rotates and diffuses with thermal evolution, causing the time-frequency representation obtained by the fractional Fourier transform to be a superposition projection of information from multiple transient evolution stages, rather than possessing a clear characteristic pointing to a single physical defect state. This feature ambiguity caused by the nature of physical evolution is difficult to overcome by extending the detection time or adjusting traditional processing procedures, severely limiting the reliability and accuracy of defect detection in engineering scenarios where the thermal history is not yet stable.
[0004] Therefore, there is an urgent need for a defect detection method that can adapt to unsteady thermal history conditions, in order to solve the technical problems of unstable defect feature representation and decreased identification reliability caused by the failure of the quasi-steady-state assumption in traditional fractional Fourier transform. Summary of the Invention
[0005] In view of this, in order to solve the problems caused by the prior art, this application provides a semiconductor defect detection method and system based on image recognition.
[0006] In a first aspect, this disclosure provides a semiconductor defect detection method based on image recognition, the method comprising:
[0007] S1. Under unsteady thermal history conditions, collect the response signals of the wafer defect region at multiple discrete moments to construct the defect response time series;
[0008] S2. Perform a fractional Fourier transform with uniform parameters on the response signal at each moment in the defect response time sequence to extract the FRFT time-frequency feature sequence that characterizes the time-frequency structure of the signal as it evolves over time.
[0009] S3. Model the FRFT time-frequency feature sequence using a convolutional long short-term memory network to generate a trajectory feature vector representing the dynamic evolution process of the defect;
[0010] S4. Based on the trajectory feature vector, determine the defect type, measure the severity, and evaluate the stability, and output the comprehensive detection result.
[0011] Optionally, S1 includes:
[0012] Establish a unified starting time reference aligned with the physical starting point of the wafer's thermal history;
[0013] Based on the starting time reference, the preset unsteady-state detection time window length, and the sampling interval, a set of discrete sampling times is generated.
[0014] At each sampling moment, optical images and auxiliary response signals are simultaneously acquired for the same defect region.
[0015] The data collected at each moment is encapsulated into a defect response time series with a clear temporal order.
[0016] Optionally, S2 includes:
[0017] The optical response intensity and auxiliary response intensity at each moment are normalized and then weighted and fused to generate a standard discrete signal of uniform length.
[0018] Based on the energy concentration index, a unified fractional Fourier transform parameter applicable to the entire unsteady evolution process is determined by scanning and selecting from the candidate orders.
[0019] Using the unified fractional Fourier transform parameters, the standard discrete signal at each time step is independently subjected to fractional Fourier transform to obtain the independent FRFT spectrum at each time step.
[0020] The total energy, main peak position, and main peak energy percentage characteristics are extracted from the FRFT spectrum at each time point and arranged in chronological order to form the FRFT time-frequency feature sequence.
[0021] Optionally, the step of scanning and selecting from candidate orders based on the energy concentration index to determine the unified fractional Fourier transform parameters applicable to the entire unsteady-state evolution process includes:
[0022] Define a candidate order scan set;
[0023] For each candidate order in the scan set, calculate its average energy concentration index on a set of representative time signals;
[0024] The candidate order that maximizes the average energy concentration index is selected as the unified fractional Fourier transform parameter.
[0025] Optionally, S3 includes:
[0026] The FRFT spectral amplitudes at each time point are energy-normalized and stacked in chronological order to construct a three-dimensional temporal feature tensor.
[0027] The temporal feature tensor is input into a convolutional long short-term memory network, and its gating mechanism is used to capture temporal dependencies and output a hidden state sequence.
[0028] The hidden state sequence is processed by exponential decay weighted averaging and segmented averaging to compress and generate a fixed-dimensional trajectory feature vector.
[0029] Optionally, during the training process of the convolutional long short-term memory network, a physical regularization constraint based on temporal continuity is introduced to make the evolution trajectory learned by the model conform to the smoothness of the physical process.
[0030] Optionally, S4 includes:
[0031] The trajectory feature vector is linearly transformed to calibrate its components to a uniform numerical range, thus obtaining a normalized feature vector;
[0032] The normalized feature vector is input into the classification model to calculate the category score, thereby determining the defect type and calculating the type confidence.
[0033] By combining the trajectory strength of the normalized feature vector, the determined defect type and its type confidence, a continuous value defect severity index is obtained through weighted calculation;
[0034] The stability of the detection process is evaluated and a stability score is generated by quantifying the fluctuations of the hidden states within the convolutional long short-term memory network at adjacent time steps.
[0035] By integrating defect type, type confidence, defect severity index, and stability score, a comprehensive detection result containing multi-dimensional information is generated.
[0036] In a second aspect, this disclosure provides a semiconductor defect detection system based on image recognition, applied to the method of the first aspect, the system comprising:
[0037] The signal acquisition module is configured to acquire the response signals of the wafer defect region at multiple discrete moments under unsteady thermal history conditions, and construct the defect response time series.
[0038] The transformation and feature extraction module is configured to perform a fractional Fourier transform with uniform parameters on the response signal at each moment in the defect response time sequence, and extract the FRFT time-frequency feature sequence that characterizes the evolution of the signal's time-frequency structure over time.
[0039] The temporal modeling module is configured to use a convolutional long short-term memory network to model the FRFT time-frequency feature sequence and generate a trajectory feature vector representing the dynamic evolution process of the defect.
[0040] The comprehensive judgment module is configured to perform defect type determination, severity measurement and stability assessment based on the trajectory feature vector, and output comprehensive detection results.
[0041] Thirdly, this disclosure provides an electronic device including a memory and at least one processor, the memory storing a computer program, and the processor executing the computer program to implement the method of the first aspect described above.
[0042] Fourthly, this disclosure provides a computer storage medium storing a computer program that, when executed, implements the method described in the first aspect.
[0043] The beneficial effects of this disclosure are that, compared with the prior art, this disclosure has the following advantages:
[0044] 1) To address the issue of continuous time-varying optical or electrical responses of defects under unsteady thermal history conditions, leading to the aliasing of different physical state information in traditional single-acquisition signals, a time reference precisely aligned with the physical starting point of thermal relaxation is constructed. High-speed synchronous acquisition is then performed at multiple moments within a clearly defined time window to obtain a time-series-clear defect response evolution sequence. This feature separates the states at different evolution moments from the data source, deconstructing the continuous transient process into a discrete but ordered observation sequence. This provides clean data that clearly distinguishes physical time evolution for subsequent analysis, fundamentally avoiding the erroneous superposition of responses from different states.
[0045] 2) To address the problems of traditional fractional Fourier transform methods, which implicitly assume quasi-steady-state stability and lead to fuzzy and unstable time-frequency features under non-steady-state conditions, a unified fractional-order parameter is determined for the entire evolution process. The signal at each time point is transformed and its features extracted independently, generating a feature sequence that can be compared temporally within the same time-frequency coordinate system. This approach avoids the feature superposition and fuzziness caused by mixing and transforming the entire time window signal. It transforms the time-frequency drift of the defect response with thermal history into a regular change in the spectral morphology under the same observation perspective, thus obtaining a time-series feature trajectory that can stably characterize the non-steady-state process.
[0046] 3) To address the challenge of assessing and ensuring the reliability of defect detection due to the uncertainty of thermal evolution, a convolutional long short-term memory network is used to learn the evolution patterns of features and output dynamic trajectory features. Based on this, a multi-dimensional judgment is made by integrating defect type, severity, and a stability score quantified based on internal state fluctuations. This feature expands the identification of objects from single static features to complete dynamic behavioral patterns, and provides a reliability measure for the judgment results through stability evaluation. Finally, a comprehensive report supporting gradient-based decision-making is output, significantly improving the reliability and decision robustness of the detection system under non-steady-state perturbation environments. Attached Figure Description
[0047] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.
[0048] Figure 1 A flowchart of a semiconductor defect detection method based on image recognition provided in an embodiment of this disclosure is shown;
[0049] Figure 2 This diagram illustrates the relationship between trajectory modeling and comprehensive judgment fusion provided in an embodiment of this disclosure.
[0050] Figure 3 A schematic diagram of the structure of a semiconductor defect detection system based on image recognition provided in an embodiment of this disclosure is shown.
[0051] The accompanying drawings have illustrated specific embodiments of this disclosure, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concepts of this disclosure to those skilled in the art through reference to particular embodiments. Detailed Implementation
[0052] The present disclosure will be further described below with reference to the accompanying drawings. The following embodiments are only used to illustrate the technical solutions of the present disclosure more clearly, and should not be used to limit the scope of protection of the present disclosure.
[0053] The components of the embodiments of the invention described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0054] In the following, the terms “comprising,” “having,” and their cognates, which may be used in various embodiments of the invention, are intended only to indicate a particular feature, number, step, operation, element, component, or combination thereof, and should not be construed as excluding, firstly, the presence of one or more other features, numbers, steps, operations, elements, components, or combinations thereof, or adding the possibility of one or more features, numbers, steps, operations, elements, components, or combinations thereof.
[0055] Unless otherwise specified, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which the various embodiments of the invention pertain. Terms (such as those defined in commonly used dictionaries) shall be interpreted as having the same meaning as in their contextual meaning in the relevant technical field and shall not be interpreted as having an idealized or overly formal meaning, unless clearly defined in the various embodiments of the invention.
[0056] In semiconductor wafer manufacturing, real-time online inspection following rapid thermal processing is crucial for controlling yield. However, at this stage, the wafer is in a non-steady-state thermal history phase, where its internal stress and electrical properties are still changing rapidly. This causes the optical or electrical response of defects to exhibit a transient process that continuously drifts over time. Traditional defect detection methods based on fractional Fourier transform rely on the quasi-steady-state assumption that the signal's time-frequency characteristics are stable within the detection time window. Under non-steady-state conditions, this assumption fails, and the time-frequency characteristics of the defect response rotate and diffuse. This results in a single transform becoming a fuzzy superposition of information from multiple evolution stages, leading to a loss of feature directivity and severely limiting the reliability of the detection.
[0057] In view of this, the present disclosure provides a semiconductor defect detection method and system based on image recognition. The aim is to construct a multi-moment response sequence synchronized with thermal history and employ unified time-frequency parameters and temporal deep learning modeling to deconstruct the continuous defect evolution process into a clear dynamic trajectory, thereby achieving stable characterization and accurate identification of defects under unsteady-state conditions and resolving the aforementioned contradictions. The implementation process of this method will be described in detail below with reference to the accompanying drawings and through several embodiments.
[0058] Figure 1 A flowchart of the semiconductor defect detection method based on image recognition provided in this disclosure is shown below. Figure 1 As shown, the process may include the following steps:
[0059] S1: Under unsteady thermal history conditions, collect the response signals of the wafer defect region at multiple discrete moments to construct the defect response time series.
[0060] This step establishes a precise temporal observation framework for wafer defect responses in the non-steady-state thermal history stage. By determining a time reference aligned with the physical starting point of thermal relaxation, high-speed synchronous acquisition at multiple time points is planned and executed. This deconstructs the originally continuously drifting and aliased transient response into a discrete evolution sequence that strictly corresponds in time and space, thereby separating different physical states from the data source and laying the foundation for subsequent analysis. This is specifically achieved through the following sub-steps.
[0061] S1.1: By receiving the process end signal from the rapid thermal processing equipment and compensating for the calibrated thermal inertia delay, a unified start time reference aligned with the physical starting point of the wafer thermal history is determined.
[0062] First, a unified and repeatable physical time starting point needs to be established for the entire unsteady-state detection process. This starting point must be closely correlated with the actual end time of the wafer's rapid thermal processing, rather than relying on unstable manual operations or equipment signal delays. Specifically, the system receives a process end signal from the rapid thermal processing equipment. This signal typically includes the nominal time for heating power shutdown and the trigger time for wafer transfer. The system records this nominal end time as... .
[0063] However, there is an inherent system delay between the time the equipment control system issues a command and the time when the wafer temperature actually begins to change significantly. To compensate for this delay and ensure the physical accuracy of the time base, a delay compensation time obtained through equipment calibration needs to be introduced. The value of this compensation time typically ranges from 10 milliseconds to 500 milliseconds, and the specific value is obtained by measuring and fitting the thermal response curve of a specific device under the same process conditions.
[0064] Unified start time reference for unsteady state detection It is calculated using the following formula: The essence of this formula is to establish an engineered, repeatable time zero point aligned with the physical starting point of the wafer's thermal history—a unified starting time reference. The key here is... It compensates for the lag in temperature changes caused by thermal inertia, rather than simply electrical or command delays. By quantifying this delay into a parameter calibrated by the thermal response curve, it ensures... Aligning with the physical moment when the wafer begins free cooling (or thermal relaxation) ensures that all subsequent sampling is bound to a unified thermal evolution phase, avoiding evolution sequence distortion caused by ambiguity in the physical meaning of the starting point. All subsequent signal acquisition actions will be based on this. This serves as a reference, ensuring the consistency of time series between different wafers and different testing batches, and eliminating thermal history deviations caused by starting point drift.
[0065] S1.2: Based on the starting time reference, the preset unsteady-state detection time window length and sampling interval, calculate and construct a set of discrete sampling moments covering key evolution stages.
[0066] In obtaining an accurate start time reference Next, it is necessary to determine the time range for observation and the specific sampling points within that range. This process constructs a discretized time observation framework. First, based on the material properties of the target wafer, the specific semiconductor process node, and empirical data from rapid thermal processing, a total time window length for unsteady-state detection is set. This window covers the critical stages in which the defect response undergoes significant evolution, and its length... The value is typically set between 10 milliseconds and 500 milliseconds.
[0067] Next, within this time window, a series of discrete sampling times need to be planned. To obtain uniform temporal evolution information, equally spaced sampling is usually adopted. The time interval between two adjacent samples is set to... . The choice of needs to strike a balance between the signal change rate and the system's acquisition capability; its value typically ranges from 0.5 milliseconds to 20 milliseconds. Based on , and This allows us to calculate all sampling times.
[0068] The i-th sampling time From the formula Given, where i is a positive integer representing the sampling sequence number. The total number of sampling times N is determined by the total time window length and the sampling interval, ensuring that the last sampling time is no later than […]. Satisfying the formula ,in This represents the floor operation. Through the above calculations, a definite set of discrete-time sampling moments is finally obtained. This set provides strict time control instructions for subsequent synchronous acquisition of defect responses at precise time points, thereby ensuring that the complete process of defect response evolution over time can be captured, rather than compressing it into a single aliased signal.
[0069] S1.3: At each sampling moment of the construction, the average optical response intensity and the average auxiliary response intensity within a short integration time window are simultaneously acquired for the same defect region on the wafer.
[0070] Based on the established set of discrete-time sampling moments, the system at each specified moment... This involves synchronous data acquisition of the same spatial defect region pre-positioned or scanned and identified on the wafer. The goal of the acquisition is to simultaneously obtain data about this region over time. The two types of response information—optical image response and auxiliary response signal—form a pair of data that are strictly corresponding in time and space.
[0071] In terms of optical image response acquisition, the defect area is acquired through an optical imaging system. Image at that moment. The defect area consists of a set of pixels. Definition. To obtain a stable and representative intensity value and avoid single-pixel noise interference, the average optical response intensity of all pixels in the region is typically calculated. The calculation formula is as follows: Where A represents the defect area. The total number of pixels contained within. Indicates at time The grayscale value or light intensity value measured at pixel position p.
[0072] In terms of auxiliary response signal acquisition, other physical signals from the same defect region are acquired simultaneously, such as reflectance spectral intensity, transient thermal radiation, or micro-area electrical response, and are denoted as continuous-time signals. In order to match the discrete sampling time To match and obtain the stable signal strength near that moment, it is necessary to... In A short time window centered on The integral averaging is performed within the time window. This measurement integration time window... The value range is generally from 0.1 milliseconds to 5 milliseconds. (Time interval) auxiliary response intensity From the formula Calculated.
[0073] Through the above synchronous acquisition, for each sampling time They all obtained a corresponding data pair. By iterating through all sampling times sequentially, a set of multi-time defect response data pairs arranged in chronological order is finally obtained. This process ensures that the differences in data collected at different time points mainly stem from the evolution of the defect response itself over thermal history, rather than noise introduced by spatial location offsets or inconsistent sampling methods.
[0074] S1.4: Assign time-series weights to the optical and auxiliary response data collected at each sampling time and encapsulate them into a structured defect response time series with a clear time order.
[0075] After collecting data at all points in time, these discrete data pairs need to be integrated into a structured sequence with a clear temporal order for subsequent processing. This encapsulation operation is not a simple data stacking, but rather strengthens its temporal order attribute through encoding.
[0076] Optical response intensity collected at each time point With auxiliary response intensity Combine them and assign a weight label related to their time sequence to each moment. Weight The design aims to preserve time order within the data structure, for example, by making it monotonically increasing with sequence number i. Subsequently, all information within the entire non-steady-state time window is encapsulated into a complete defect response time series D. The structured representation of this series can be described in the following form: .
[0077] This step ultimately forms a complete and traceable sequence of nonsteady-state defect responses. This sequence clearly distinguishes the physical states at different evolutionary moments from the data source, laying a solid foundation for subsequent analysis and fundamentally avoiding the erroneous compression or characterization of the entire thermal evolution process as a single steady-state feature.
[0078] In the technical solution of this disclosure, by constructing a time reference precisely aligned with the physical starting point of the wafer's thermal history and performing multi-moment synchronous signal acquisition within a defined non-steady-state time window, the problem of fuzzy and aliased defect response signals during the drastic thermal change phase of the wafer after rapid thermal processing is fundamentally solved. This step deconstructs the originally continuously drifting transient response process into a defect response evolution sequence arranged in strict temporal order and corresponding in spatial location, providing a clean and time-defined data foundation for subsequent analysis. By compensating for thermal inertia delay and setting sampling times with clear physical meaning, the consistency of time series between different batches of tests is ensured, eliminating errors introduced by starting point drift or asynchronous sampling. This ensures that the differences in the acquired data mainly stem from the evolution of the defect response itself, rather than measurement noise or operating condition fluctuations.
[0079] S2: Perform a fractional Fourier transform with uniform parameters on the response signal at each moment in the defect response time sequence to extract the FRFT time-frequency feature sequence that characterizes the evolution of the signal's time-frequency structure over time.
[0080] This step aims to transform the time-series defect response sequence into a set of comparable time-frequency features. It establishes a unified fractional Fourier transform parameter for the entire non-steady-state process and uses this parameter to independently transform and extract features from the signal at each time step. This is equivalent to transforming the continuous drift of the defect response into a regular temporal change in spectral morphology from a fixed time-frequency perspective, thereby generating a structured time-series time-frequency feature sequence. This is specifically achieved through the following sub-steps.
[0081] S2.1: Normalize and weight the optical response intensity and auxiliary response intensity at each moment to generate a standard discrete signal of uniform length, which constitutes a standard input signal set.
[0082] Based on the defect response sequence output in step S1, i.e., the set of data pairs sorted by time. This requires converting the multimodal response data at each time step into a standard discrete signal form suitable for fractional Fourier transform (FRFT). This operation ensures that signals at different time points have the same length and comparable dimensions, laying the foundation for the subsequent application of unified FRFT parameters.
[0083] For each time i, its optical response intensity With auxiliary response intensity Normalization and weighted fusion are performed to generate a standard discrete signal of length L. , where n is the discrete sample index, ranging from 0 to L-1. The length L of the discrete signal is a constant, set according to a trade-off between system sampling capability and computational complexity, typically ranging from 64 to 2048.
[0084] In practice, a normalized weighted combination method is used. Optical response Divide by a normalization coefficient This coefficient is usually taken from all wafers in the same batch. Statistical measures of the value, such as its 95th percentile or maximum value, are used to avoid extreme values dominating the signal. Similarly, auxiliary responses... Divide by its corresponding normalization coefficient Subsequently, the two are weighted by a modal weighting coefficient determined by engineering calibration. Combine them. This process is described by the formula. Description. Among them, the modal weighting coefficients... It is a constant, ranging from 0.10 to 2.00, used to balance the contributions of different modal signals in the final characteristics.
[0085] Through the above processing, a standardized discrete input signal is obtained for each time i. After iterating through all N time points, a set of standard input signals is obtained. The essence of this step is to merge the multiple physical responses of the same defect at the same time into a signal representation under a unified numerical system, thereby ensuring that the input form is consistent and comparable when performing FRFT transformation independently at each time point.
[0086] S2.2: Based on the energy concentration index, a unified fractional Fourier transform parameter applicable to the entire unsteady evolution process is determined by scanning and selecting from the candidate orders.
[0087] To ensure that the time-frequency representations obtained after FRFT transformation at different times lie in the same time-frequency rotating coordinate system, thus enabling direct comparison, a unified fractional-order parameter must be determined and used for all times in the entire sequence. Here, the candidate parameter for scan evaluation is denoted as... The final determined optimal unified parameters are denoted as If different orders are adaptively selected for each time step, the FRFT results at each time step will be projected onto different time-frequency planes, resulting in a loss of comparability.
[0088] First, establish fractional-order candidate parameters. Its corresponding time-frequency plane rotation angle The relationship is given by the formula. Given. Among them, The value range is usually kept away from degenerate cases such as 0 and 2, and in engineering, it is often taken to be between 0.10 and 1.99.
[0089] To select the optimal unified order Define and adopt an evaluation index based on energy concentration. From all N times, select a representative subset Q of times, for example, selecting 10 to 50 times at fixed time intervals. For each candidate... Value, calculate the signal at each time step in subset Q. of The spectrum is obtained by performing an FRFT transformation. , where u is the spectral index. Then, the candidate is calculated. Average energy concentration index on subset Q .
[0090] Energy Concentration Index The definition of is: The larger the value of this index, the more concentrated the signal energy is in the FRFT domain, the sharper the spectrum, which usually means that the defect response characteristics are more obvious.
[0091] In the engineering implementation, a candidate order scan set is defined. Each element, for example to Both represent candidate order; scan step size A value between 0.02 and 0.10 is recommended to balance accuracy and computational efficiency. Iterate through and compute each candidate... Corresponding average concentration Ultimately, it led to That is, the order that maximizes the average energy concentration is selected as the parameter for the unified fractional Fourier transform. And according to the formula... Obtain the corresponding uniform rotation angle This step establishes a unified time-frequency observation perspective effective for the entire unsteady evolution process through physically meaningful indicators. Under unsteady conditions, if an optimal fractional order is adaptively selected for each time step, the FRFT spectra at each time step will be projected onto different time-frequency planes, causing their spectral characteristics (such as the position of the main peak) to lose direct temporal comparability. This method enforces the use of a unified... This is equivalent to fixing a time-frequency observation lens. Any energy diffusion or rotation of the defect response over time will be manifested as a continuous change in the spectral morphology under this fixed lens, thus transforming the time-varying physical process into an analyzable time-series characteristic trajectory, which is a prerequisite for subsequent time-series modeling.
[0092] Understandably, the significance of unifying the fractional Fourier transform parameters lies in fixing the same time-frequency observation perspective for the entire unsteady evolution process, so that the fractional Fourier transform results of the defect response at each discrete sampling time are in the same time-frequency rotating coordinate system, thereby ensuring that the FRFT spectrum and its derived features at each time have direct temporal comparability, and avoiding the loss of consistent reference for spectral features such as the main peak position and energy distribution due to different transformation perspectives at each time.
[0093] S2.3: Using the unified fractional-order parameters, perform fractional-order Fourier transforms independently on the standard discrete signal at each time step to obtain independent FRFT spectra at each time step.
[0094] Determining the uniform fractional order parameter Then, each standard discrete signal in the standard input signal set generated in step S2.1 can be processed. independently The fractional Fourier transform of order 1. This step strictly prohibits any form of cross-time weighting, averaging, or fusion of the transform results at different times, in order to fully preserve the original time-frequency characteristics of the response at each time point.
[0095] For the i-th time, its Spectrum obtained by FRFT transformation The formula for calculating Discrete FRFT is given as follows: .in, express The discrete kernel function of the fractional Fourier transform is specifically determined by the device sampling interval, signal length L, and order. The decision is made jointly. In engineering implementation, the mature Fast Fractional Fourier Transform algorithm can be used for calculation.
[0096] The above transformation is performed sequentially on i from 1 to N, ultimately obtaining the independent FRFT spectrum set at each time step. The essence of this step is a fixed rotation angle. Within the defined unified time-frequency coordinate system, the projection of the defect response signal at each moment is observed. This is fundamentally different from traditional methods that treat the entire time window signal as a whole for FRFT transformation under the quasi-steady-state assumption. The transformation result of the traditional method is essentially a superposition and ambiguity of features from multiple evolution stages. However, this step, through a unified perspective and independent transformation, decouples the continuous evolution of the defect with its thermal history into a series of clear and time-comparable time-frequency snapshots, laying the foundation for subsequent accurate modeling of the evolution trajectory.
[0097] Thus, the time drift of the defect response due to thermal history evolution is transformed into its FRFT spectrum. The shape, peak position, and energy distribution of the energy change over time i, and this change is clear and traceable.
[0098] S2.4: Extract the total energy, main peak position, and main peak energy ratio from the FRFT spectrum at each time step, and encapsulate them into an FRFT time-frequency feature sequence in chronological order.
[0099] To facilitate subsequent temporal deep learning network processing and further highlight core features, it is necessary to obtain the complete FRFT spectrum from each time step. In this process, several scalar features with clear physical meaning are extracted and encapsulated into a feature sequence in chronological order.
[0100] Typically extracted features include: the total energy of the FRFT spectrum at that moment. The location of the main energy peak And the proportion of the main peak energy to the total energy. These features reflect the overall intensity of the defect response at that moment, the main projection position in the unified time-frequency coordinate system, and the sharpness and stability of the features, respectively.
[0101] The specific calculation methods for each feature are as follows:
[0102] Total Energy From the formula calculate.
[0103] Main peak location It is obtained by finding the index corresponding to the maximum value of the energy spectrum, i.e. .
[0104] Main peak energy percentage From the formula Calculation. Among them, For example, a very small constant arrive This is used to prevent the denominator from being zero.
[0105] Subsequently, these features at time i are combined into a feature vector. Arrange all [items] in chronological order from i=1 to N. The arrangement, i.e., the formation of the required FRFT time-frequency feature sequence. This sequence, as the final output of step S2, essentially extracts the continuous changes in the FRFT spectrum during the non-steady-state evolution into a series of time-evolving feature trajectories that can be directly processed by mathematical models, providing accurate input for the next step of time series modeling.
[0106] In the technical solution of this disclosure, by determining a unified fractional Fourier transform parameter for the entire unsteady evolution process and independently transforming the response signal at each time step using this unified parameter, the continuous time-frequency drift of the defect response with thermal history is successfully transformed into a series of spectral feature sequences that can be clearly compared under the same time-frequency coordinate system. This method strictly avoids the traditional practice of mixing and superimposing time-frequency information from different evolution stages, thereby preventing feature ambiguity. By extracting features such as the total energy, main peak position, and energy percentage of the FRFT spectrum at each time step and encapsulating them into a time-series sequence, the complex unsteady evolution process is refined into a structured feature trajectory that can be directly processed by mathematical models, laying a precise input foundation for subsequent time-series dynamic modeling.
[0107] S3: The FRFT time-frequency feature sequence is modeled using a convolutional long short-term memory network to generate a trajectory feature vector representing the dynamic evolution process of the defect.
[0108] This step utilizes a convolutional long short-term memory network to learn and represent the dynamic evolution of defects from the FRFT time-frequency feature sequence. By constructing the temporal features as tensor inputs to the convolutional long short-term memory network, the model can capture the temporal dependencies between features, and then fuse and compress discrete feature points to generate a fixed-dimensional trajectory feature vector, thereby completing the transformation from capturing instantaneous states to representing the dynamic behavior of the entire process. Figure 2 The trajectory modeling and comprehensive judgment fusion relationship diagram provided in the embodiments of this disclosure, such as Figure 2 As shown, this is achieved through the following sub-steps.
[0109] S3.1: After energy normalization of the FRFT spectral amplitudes at each time step, stack them in chronological order to construct a three-dimensional temporal feature tensor that can be used as input to a convolutional long short-term memory network.
[0110] To model using a convolutional long short-term memory network, the FRFT time-frequency feature sequence output in step S2 first needs to be processed. Organize it into a tensor form acceptable to the network. Each of these tensors... At least includes total energy Main peak location and the proportion of main peak energy Three scalar features. The goal of this step is to construct an input tensor that preserves both temporal order and the local structure information of the FRFT spectrum at each time step.
[0111] To achieve this goal, it is preferable to use the complete FRFT spectral amplitude sequence at each time step. As the foundational data, to eliminate the impact of overall energy differences between different time points on model learning, the spectrum at each time point needs to be normalized. Specifically, for the i-th time point, its normalized spectral amplitude bands... Through formula Calculated. Here, It is the total energy of the spectrum at that time. It is a very small constant, with a range of values of 1000. arrive This is used to ensure numerical stability. The essence of the normalization operation is to transform the spectral shape at each time step into a relative energy distribution, so that subsequent models focus more on the evolution of the spectral shape over time, rather than changes in absolute intensity.
[0112] Subsequently, the normalized spectral amplitude bands at all N time points were analyzed. Stacked sequentially over time, these slices form a three-dimensional temporal feature tensor Z. The temporal dimension of this tensor is N, and the spatial dimension (or feature dimension) is L, which is the length of the FRFT spectrum. Thus, each slice in tensor Z represents the normalized projection of the defect response at a given moment onto a unified FRFT coordinate system, and the entire tensor constitutes a three-dimensional data block describing how the projection evolves over time. If the complete spectrum is not preserved in the engineering implementation... Then the triplet features need to be... While extending the rules into fixed-length feature vectors can simulate local structures, using the full spectral magnitude is a better choice.
[0113] Understandably, the significance of the temporal feature tensor lies in stacking the FRFT spectral amplitudes obtained in step S2 in chronological order after energy normalization, forming a unified input expression that simultaneously contains temporal order and spectral local structure information, thereby satisfying the dual requirements of convolutional long short-term memory networks for temporal modeling and local structure extraction; this tensor enables the network to continuously capture the evolution of FRFT spectral morphology driven by non-steady-state thermal history at each time step based on the current input and historical hidden state.
[0114] As an optional embodiment, the rule expansion can be achieved by constructing a... This is achieved using a local Gaussian distribution vector centered at the center: ,in, This is a preset width parameter. This method can simulate the local spectral structure near the main peak to some extent. However, directly using the fully normalized spectral amplitude... It can preserve more comprehensive spectral details, including possible secondary peaks or spectral width variations, which are more helpful in depicting subtle evolutionary processes.
[0115] S3.2: Input the temporal feature tensor into the convolutional long short-term memory network, capture the temporal dependencies through its gating mechanism, and output the hidden state sequence.
[0116] After obtaining the input tensor Z, it is fed into a convolutional long short-term memory network for temporal modeling. This network, through a unique gating mechanism and memory units, fuses current input and historical information at each time step, thereby capturing the dynamic evolution of the time series. The computation of the network at each time step i begins with the generation of gating signals. These gating signals control the flow of information, such as determining how much historical information is forgotten, how much new information is stored, and what content is output. (Gating pre-activation amount) The calculation is one of the core steps, and its specific form is defined by the following formula: ,in, The input feature strip represents the current time i. The hidden state outputs of the network at the previous time step i-1, together forming the basis of gating computation, are represented by parameters a, b, and c. These are the trainable parameters of the network. During the initial training phase, a and b are typically assigned small random values, such as in the range of -0.1 to 0.1, while c is usually initialized to 0. These parameters are continuously optimized through gradient descent during subsequent training, enabling the network to learn how to effectively combine the current input with historical states.
[0117] In engineering implementation, the above-described gating pre-activation calculations are used for the forget gate, input gate, and output gate, respectively. Each gate has an independent set of parameters a, b, and c, thereby finely controlling the information flow for different functions. Based on these gating signals, the network then updates its internal memory state. The memory state carries the network's comprehensive memory of sequential historical information, and its update process emphasizes the preservation of historical memory and the integration of current information. An intuitive and simplified expression can be described as follows: ,in, The memory retention coefficient is a hyperparameter or a quantity learned by the network, ranging from 0.60 to 0.98. The larger the value, the more the network tends to retain past information when updating its memory, making it more suitable for scenarios with a slow evolutionary trend; The smaller the value, the faster the network responds to newly received data. More sensitive.
[0118] Output of the final hidden state at the current time step Then it is determined by the updated memory state It is obtained through a nonlinear transformation. This transformation aims to normalize the range of the output value and introduce nonlinearity. An exemplary calculation formula is as follows: The formula is used to calculate... This refers to the hidden state output by the network at the i-th time step. This operation will memorize the state. The magnitude of the gradient is compressed to approximately the range (-1, 1), which helps maintain numerical stability during training and inference, preventing gradient explosion or vanishing. After processing all N time steps in sequence, the hidden states at each time step are obtained. Hidden state sequence constructed in chronological order This is the output of the sub-step.
[0119] S3.3: Perform exponential decay weighted averaging and piecewise averaging on the hidden state sequence output by the network to compress and generate a fixed-dimensional defect evolution trajectory feature vector.
[0120] The hidden state sequence fully records the network's internal understanding of temporal features, but its dimensionality still changes with time steps. In order to obtain a fixed-dimensional feature vector that can represent the entire evolution process for final judgment, the sequence needs to be compressed and converged.
[0121] Considering that the defect response changes are often most dramatic and information-rich in the early stages of unsteady thermal history, different weights can be assigned to different moments during compression to emphasize the early evolution stages. An effective weighting method is to use exponential decay, i.e., the weight at moment i is... From the formula Calculation. Among them, It is the time decay constant, with a value ranging from 3 to 50. The smaller the value, the faster the weight decays, and the more prominent the emphasis is on the early moments of the sequence.
[0122] By applying a weighted average to the hidden state sequence, preliminary trajectory features can be obtained. The calculation formula is: .this It is a vector of length L, which can be regarded as a summary of the defect response evolution trajectory in the spectral domain during the entire unsteady time period.
[0123] To further reduce dimensionality and enhance feature robustness, one can... Perform piecewise averaging. Divide the vector of length L into M segments, where M is a constant, typically between 8 and 128. Calculate all values within each segment. The average value is used as the characteristic value of this segment. ,Right now ,in This represents the length of each segment. Ultimately, we obtain an M-dimensional defect evolution trajectory feature vector Y. This vector, in a fixed and compact form, encapsulates the core dynamic behavior pattern of the defect response throughout the entire thermal evolution process.
[0124] The physical meaning of the evolutionary trajectory feature vector Y lies in the fact that each of its components corresponds to the average performance of the energy convergence of a specific sub-interval in the unified time-frequency domain of FRFT over time. For example, if the time-frequency peak position of the response energy of a defect undergoes a systematic shift during the cooling process, the components in the Y vector corresponding to different sub-intervals will exhibit specific pattern differences. Therefore, Y is essentially a compressed encoding of the dynamic fingerprint of a defect, and its adaptability to non-steady-state conditions stems from the generalization of the complete evolutionary process, rather than the capture of a certain instantaneous state.
[0125] S3.4: Introduce physical regularization constraints based on time continuity during model training, and use the trained network to infer new sequences to output the final evolutionary trajectory feature vector.
[0126] To ensure that the model learns physically plausible evolutionary patterns during training and maintains stable output during online inference, specific constraints need to be imposed on the training process. During the training phase, a large number of labeled defect response sequence samples are used, with the trajectory feature vector Y as a bridge. Classification loss functions, such as cross-entropy loss, are used to optimize the network parameters, enabling it to distinguish different types of defect evolutionary patterns.
[0127] Simultaneously, a time continuity regularization term based on physical priors is introduced. This regularization term encourages the hiding of states. The changes between adjacent time steps are smooth, which aligns with the common sense that thermal history-driven physical evolution typically exhibits continuity. The continuity constraint term R is defined as follows: The total loss function during training, Loss, is the classification loss. The weighted sum of this regularization term R, i.e. ,in This is a regularization weight, ranging from 0.001 to 0.50, used to balance classification accuracy and trajectory smoothness. The introduction of this continuous regularization term R is based on the physical prior that the evolution of the defect response is driven by the thermal history and should be smooth over a short timescale. It is not simply mathematical smoothing, but rather forces the model to learn an evolutionary pattern that conforms to physical laws, thereby improving the model's robustness to noise interference and enhancing the characteristics of its output evolutionary trajectory. It better reflects the real physical process, rather than the fluctuations caused by measurement noise.
[0128] During the online inference phase, the pre-trained network parameters are fixed, and a new defect response sequence is input. The network will then go through steps S3.1 to S3.3, ultimately outputting a corresponding evolutionary trajectory feature vector. Because a continuity constraint was added during training, the inference results in... It is more representative of smooth and reliable physical evolution trajectories, rather than anomalous patterns caused by measurement noise or sudden changes. This will serve as the direct basis for the comprehensive defect assessment in step S4.
[0129] In the technical solution of this disclosure, a convolutional long short-term memory network is used to model the FRFT time-frequency feature sequence. The core of this approach is to learn and characterize the complete evolution of the defect response under the entire non-steady-state thermal history, rather than extracting a single static feature. This step involves constructing the temporal features as tensors and inputting them into the network. Utilizing the network's temporal dependency modeling capability, feature points at discrete moments are connected into a smooth physical evolution trajectory. By introducing time continuity constraints based on physical priors during training, the evolution trajectory features output by the network better reflect the real physical process and suppress noise interference. The resulting fixed-dimensional evolution trajectory feature vector is a compressed encoding of the defect's dynamic behavior pattern, realizing a shift from capturing instantaneous states to representing the entire dynamic process, significantly improving the model's adaptability and generalization ability to non-steady-state conditions.
[0130] S4: Based on the trajectory feature vector, perform defect type determination, severity measurement and stability assessment, and output comprehensive detection results.
[0131] This step, based on evolutionary trajectory feature vectors, performs a final comprehensive judgment and output on defects. After eliminating scale differences through normalization, the system sequentially performs defect type determination and severity quantification, and evaluates the stability of the detection process itself by quantifying the internal state fluctuations of the model. Finally, it integrates type, confidence, severity, and stability scores to output a comprehensive report supporting tiered production line decisions. Figure 2 As shown, this is achieved through the following sub-steps.
[0132] S4.1: Perform a linear transformation on the evolution trajectory feature vector to calibrate each component to a uniform numerical range, thereby obtaining a normalized feature vector to eliminate scale differences.
[0133] The defect evolution trajectory feature vector output in step S3 The eigenvector encapsulates the dynamic behavior of defects throughout the entire thermal evolution process. However, different wafer batches, different process conditions, or different thermal histories may lead to differences in the overall amplitude level of this eigenvector. To ensure the universality and robustness of subsequent decision-making logic, the eigenvector needs to be calibrated to normalize it to a uniform numerical range.
[0134] set up Let M-dimensional vector be represented as First, calculate the minimum value among all components of the vector. and maximum value Then, a linear transformation is performed on each component using the following formula: .in, It is the m-th characteristic component after calibration. It is a very small positive number, with a value range of 1. arrive Its function is to prevent the denominator from being zero and to ensure the stability of numerical calculations.
[0135] After the above operations, the original feature vector Converted into a new vector All element values of the vector are compressed into a closed interval between 0 and 1. The essence of this step is to eliminate the feature scale differences caused by fluctuations in overall signal strength, so that subsequent defect type classifiers, severity calculation modules, etc., can operate based on a unified, dimensionless feature space, thereby significantly improving the adaptability of the judgment system under different production batches and operating conditions.
[0136] S4.2: Input the normalized feature vector into the classification model to calculate the category score, thereby determining the defect type and calculating the confidence level of the determination.
[0137] Based on the normalized feature vector V, the goal of this sub-step is to determine the specific category of the defect and to provide the confidence level of that determination. This is achieved through a lightweight classification model that maps high-dimensional trajectory features to discrete defect types.
[0138] Assume the total number of known defect categories is K, and the category set is... For each defect category k, a corresponding score can be calculated. The score is composed of a linearly weighted combination of the feature vectors V plus a bias term: .
[0139] in, These are the weight coefficients corresponding to the k-th defect type and the m-th feature component in the classification model. These are the bias constants for the k-th type of defect. and During the training phase of the model, a large amount of sample data labeled with known defect types is used to learn the model through an optimization algorithm.
[0140] The predicted defect type C was determined to receive the highest score. That category: .
[0141] To quantify the reliability of this judgment, a type confidence score is calculated. The scores are converted into probabilistic form using exponential normalization. .
[0142] here, It predicts the score corresponding to category C. It is an exponential function. The value range is (0,1), and the closer its value is to 1, the greater the confidence of the classification model in classifying the current sample as type C. This confidence level not only provides users with a reference for the reliability of the judgment, but also participates in subsequent severity calculations and comprehensive decision-making.
[0143] S4.3: Combining the intensity of trajectory features, the identified defect types and their confidence levels, a continuous defect severity index is obtained through calculation and coefficient weighting.
[0144] Defect severity assessment requires comprehensive consideration of its type and the intensity characteristics of its evolution trajectory. This sub-step calculates a continuous defect severity index based on the normalized feature V and the identified defect type C. This indicator reflects the potential harm that defects may cause to device performance and reliability.
[0145] First, a trajectory intensity term is constructed from the perspective of evolutionary trajectory. This index is characterized by the mean of the squares of the components of the eigenvector V, and its calculation formula is as follows: The squaring operation enhances the contribution of salient feature components while avoiding information loss that might result from the cancellation of positive and negative values of different feature components. Because... , The value range is also between 0 and 1.
[0146] Secondly, a severity coefficient related to the defect type is introduced. This coefficient is a predefined constant, its value determined based on process knowledge, historical failure analysis data, or statistical analysis of its impact on yield. It is used to characterize the differences in the inherent severity of different types of defects. For example, the severity of certain fatal defects... The value will be significantly higher than that of a general defect, with a typical range between 0.5 and 5.0.
[0147] Ultimately, the overall trajectory intensity Type coefficient and type confidence Calculate the continuous value of defect severity : .in, The item serves as a moderating factor, incorporating the confidence level of the type determination into the severity assessment. When the confidence level... When the level is low, the factor will decrease accordingly. The value reflects a conservative assessment strategy, avoiding overestimation of severity when the type is uncertain. The higher the value, the more severe the defect. Based on the actual needs of the production line, a series of thresholds can be set to... It is further divided into discrete severity levels.
[0148] S4.4: Quantify the fluctuations in the internal state of the convolutional long short-term memory network to assess the stability of the detection process, and integrate the type, confidence, severity and stability scores to generate a comprehensive detection result.
[0149] When conducting tests under unsteady thermal histories, the stability of the judgment results is crucial. Stability assessment aims to answer the question: under the current thermal perturbation, is the observed defect evolution trajectory smooth and continuous, thus ensuring that the judgment based on this trajectory is reliable and repeatable?
[0150] An effective method for stability assessment is to quantify the degree of volatility during the evolution process. This is especially true if the hidden state sequence generated by the convolutional long short-term memory network in step S3 is accessible. Instability can then be defined by calculating the mean square value of the difference between the hidden states in adjacent time steps. : Where N is the total number of time steps, and L is the length of the hidden state strip. The larger the value, the more drastic the changes in the hidden state over time, and the more unstable the corresponding defect evolution trajectory.
[0151] Instability Converted into a more intuitive stability score Its value range is usually between 0 and 1, with a larger value indicating greater stability. The conversion formula can be expressed as: .in, It is a scaling constant used to... The numerical range is mapped to a suitable interval, and its specific value needs to be calibrated based on training data or engineering experience. A suggested range is within [specific range]. arrive between.
[0152] Finally, the system integrates all intermediate results to generate a final comprehensive report on wafer defect detection. This report should include at least the following core elements: defect type C, type confidence level. Defect severity Stability score Among them, the stability score It's not a direct measurement of the defect itself, but rather an evaluation of the reliability of the evolution trajectory obtained during the detection process under a non-steady-state environment. It quantifies the internal state fluctuations of the time-series model, providing an important metacognitive indicator for the final judgment. Based on this information, the system can also automatically provide handling suggestions. For example, the following rule can be set: when... and When the judgment result is deemed highly reliable, the wafer can proceed to the next process; when or In such cases, it is recommended that the wafer be re-inspected to confirm the results; when Exceeding the preset scrap threshold At that time, regardless of other indicators, it is marked as a high-risk defect and scrapping is recommended.
[0153] In the technical solution of this disclosure, a comprehensive judgment is made based on evolutionary trajectory characteristics. By integrating multi-dimensional information such as defect type, severity, and detection stability, the limitation of unstable single-moment feature judgment under non-steady-state conditions is overcome. This step first standardizes the trajectory features to eliminate batch-to-batch differences. Then, a classification model is used to determine the defect type and assign confidence levels. Simultaneously, the severity of the defect is quantified by combining trajectory strength and prior knowledge of the type. Specifically, the stability of the evolutionary trajectory obtained in this detection is evaluated by quantifying the fluctuations in the internal state of the time-series model, providing a key reliability meta-indicator for the judgment result. The final system output includes a comprehensive report containing type, confidence level, severity, and stability score, supporting tiered production line decisions from automatic release and re-inspection to scrap recommendations. This greatly enhances the reliability and operability of defect detection decisions under thermal disturbance environments.
[0154] This disclosure also provides a semiconductor defect detection system based on image recognition. Figure 3 This is a schematic diagram of a semiconductor defect detection system based on image recognition, provided according to an embodiment of this disclosure. This system is used to run the semiconductor defect detection methods based on image recognition described in the above embodiments. (Refer to...) Figure 3 The system may include:
[0155] The signal acquisition module is configured to acquire the response signals of the wafer defect region at multiple discrete moments under unsteady thermal history conditions, and construct the defect response time series.
[0156] The transformation and feature extraction module is configured to perform a fractional Fourier transform with uniform parameters on the response signal at each moment in the defect response time sequence, and extract the FRFT time-frequency feature sequence that characterizes the evolution of the signal's time-frequency structure over time.
[0157] The temporal modeling module is configured to use a convolutional long short-term memory network to model the FRFT time-frequency feature sequence and generate a trajectory feature vector representing the dynamic evolution process of the defect.
[0158] The comprehensive judgment module is configured to perform defect type determination, severity measurement and stability assessment based on the trajectory feature vector, and output comprehensive detection results.
[0159] According to embodiments of this disclosure, an electronic device is also provided, which may include a processor, a communications interface, a memory, and a communication bus, wherein the processor, the communications interface, and the memory communicate with each other via the communication bus. The processor can invoke logical instructions stored in the memory to execute the methods provided in the above embodiments.
[0160] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0161] On the other hand, this disclosure also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the methods provided in the above embodiments.
[0162] The device embodiments described above are merely illustrative. 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 modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.
[0163] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.
[0164] It should be understood that the above embodiments are only used to illustrate the technical solutions of this disclosure, and not to limit them; although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure.
Claims
1. A semiconductor defect detection method based on image recognition, characterized in that, The method includes: S1. Under unsteady thermal history conditions, collect the response signals of the wafer defect region at multiple discrete moments to construct the defect response time series; S2. Perform a fractional Fourier transform with uniform parameters on the response signal at each moment in the defect response time sequence to extract the FRFT time-frequency feature sequence that characterizes the time-frequency structure of the signal as it evolves over time. S3. Model the FRFT time-frequency feature sequence using a convolutional long short-term memory network to generate a trajectory feature vector representing the dynamic evolution process of the defect; S4. Based on the trajectory feature vector, determine the defect type, measure the severity, and evaluate the stability, and output the comprehensive detection result.
2. The semiconductor defect detection method based on image recognition according to claim 1, characterized in that, S1 includes: Establish a unified starting time reference aligned with the physical starting point of the wafer's thermal history; Based on the starting time reference, the preset unsteady-state detection time window length, and the sampling interval, a set of discrete sampling times is generated. At each sampling moment, optical images and auxiliary response signals are simultaneously acquired for the same defect region. The data collected at each moment is encapsulated into a defect response time series with a clear temporal order.
3. The semiconductor defect detection method based on image recognition according to claim 1, characterized in that, S2 includes: The optical response intensity and auxiliary response intensity at each moment are normalized and then weighted and fused to generate a standard discrete signal of uniform length. Based on the energy concentration index, a unified fractional Fourier transform parameter applicable to the entire unsteady evolution process is determined by scanning and selecting from the candidate orders. Using the unified fractional Fourier transform parameters, the standard discrete signal at each time step is independently subjected to fractional Fourier transform to obtain the independent FRFT spectrum at each time step. The total energy, main peak position, and main peak energy percentage characteristics are extracted from the FRFT spectrum at each time point and arranged in chronological order to form the FRFT time-frequency feature sequence.
4. The semiconductor defect detection method based on image recognition according to claim 3, characterized in that, The process of selecting uniform fractional Fourier transform parameters applicable to the entire unsteady-state evolution process by scanning from candidate orders based on energy concentration indices includes: Define a candidate order scan set; For each candidate order in the scan set, calculate its average energy concentration index on a set of representative time signals; The candidate order that maximizes the average energy concentration index is selected as the unified fractional Fourier transform parameter.
5. The semiconductor defect detection method based on image recognition according to claim 1, characterized in that, S3 includes: The FRFT spectral amplitudes at each time point are energy-normalized and stacked in chronological order to construct a three-dimensional temporal feature tensor. The temporal feature tensor is input into a convolutional long short-term memory network, and its gating mechanism is used to capture temporal dependencies and output a hidden state sequence. The hidden state sequence is processed by exponential decay weighted averaging and segmented averaging to compress and generate a fixed-dimensional trajectory feature vector.
6. The semiconductor defect detection method based on image recognition according to claim 5, characterized in that, During the training process of the convolutional long short-term memory network, a physical regularization constraint based on temporal continuity is introduced to make the evolution trajectory learned by the model conform to the smoothness of the physical process.
7. The semiconductor defect detection method based on image recognition according to claim 1, characterized in that, S4 includes: The trajectory feature vector is linearly transformed to calibrate its components to a uniform numerical range, thus obtaining a normalized feature vector; The normalized feature vector is input into the classification model to calculate the category score, thereby determining the defect type and calculating the type confidence. By combining the trajectory strength of the normalized feature vector, the determined defect type and its type confidence, a continuous value defect severity index is obtained through weighted calculation; The stability of the detection process is evaluated and a stability score is generated by quantifying the fluctuations of the hidden states within the convolutional long short-term memory network at adjacent time steps. By integrating defect type, type confidence, defect severity index, and stability score, a comprehensive detection result containing multi-dimensional information is generated.
8. A semiconductor defect detection system based on image recognition, applied to the method according to any one of claims 1-7, characterized in that, The system includes: The signal acquisition module is configured to acquire the response signals of the wafer defect region at multiple discrete moments under unsteady thermal history conditions, and construct the defect response time series. The transformation and feature extraction module is configured to perform a fractional Fourier transform with uniform parameters on the response signal at each moment in the defect response time sequence, and extract the FRFT time-frequency feature sequence that characterizes the evolution of the signal's time-frequency structure over time. The temporal modeling module is configured to use a convolutional long short-term memory network to model the FRFT time-frequency feature sequence and generate a trajectory feature vector representing the dynamic evolution process of the defect. The comprehensive judgment module is configured to perform defect type determination, severity measurement and stability assessment based on the trajectory feature vector, and output comprehensive detection results.
9. An electronic device, characterized in that, The electronic device includes a memory and at least one processor, the memory storing a computer program, and the processor executing the computer program to implement the image recognition-based semiconductor defect detection method according to any one of claims 1-7.
10. A computer storage medium, characterized in that, It stores a computer program, which, when executed, implements the semiconductor defect detection method based on image recognition according to any one of claims 1-7.