A high-speed camera and detection method for imageless dark-field detection of semiconductors

Abstract

This application discloses a high-speed camera and detection method for imageless dark-field inspection of semiconductors, relating to the field of semiconductor inspection technology. The method includes: integrating an adjustable laser source array, a high-performance TDI camera, and a pre-scanning sensor to collect brightness information of the pre-scanning area and generate a pre-scanning reflectivity map; using the adjustable laser source array to sequentially illuminate the pre-scanning area with different illumination lights to generate multiple scan images and extract features to generate a multimodal differential feature field; using a multimodal illumination and defect feature co-enhancement algorithm to predict the main defect types within the area and generate corresponding illumination mode commands; using a TDI parameter dynamic modulation algorithm to calculate the optimal TDI parameters and generate TDI camera control commands; scanning the wafer according to the two commands, collecting raw image data, inputting it into a deep defect classification model, and outputting a defect information map and corresponding information. This application aims to solve the problem of balancing accuracy and efficiency in semiconductor inspection.

Description

Technical Field

[0001] This invention relates to the field of semiconductor detection technology, and in particular to a high-speed camera and detection method for imageless dark field detection of semiconductors. Background Technology

[0002] As semiconductor process nodes advance to the nanoscale, nanoscale defects on wafer surfaces (such as particles, scratches, and watermarks) have an increasingly significant impact on chip yield. Imageless dark-field inspection is a key technology for detecting these defects. However, existing technologies face a core dilemma: increasing scanning speed leads to a decrease in image signal-to-noise ratio and loss of detail; while pursuing high-resolution imaging significantly increases data acquisition and processing time, affecting production cycle time. Although traditional TDI cameras can improve the signal-to-noise ratio at certain speeds, their integration series and exposure time are fixed pre-set values, unable to adapt to the dynamic changes in reflectivity and defect types in different areas of the wafer surface, making it difficult to achieve a globally optimal solution for detection performance. Summary of the Invention

[0003] This invention provides a detection method for imageless dark field detection in semiconductors, comprising:

[0004] It integrates an adjustable laser light source array, a high-performance TDI camera, and a pre-scanning sensor. The pre-scanning sensor collects brightness information of the pre-scanning area and generates a pre-scanning reflectance map in real time.

[0005] By using an adjustable laser light source array, the pre-scanning area is illuminated sequentially with illumination light of different wavelengths and incident angles to generate multiple scan images and extract features to generate a multimodal differential feature field.

[0006] For multimodal differential feature fields, a multimodal illumination and defect feature collaborative enhancement algorithm is used to identify defects, predict the main defect types in the region, and generate corresponding illumination mode instructions based on the defect types.

[0007] Based on the pre-scan reflectance spectrum, the optimal TDI parameters are calculated using a TDI parameter dynamic modulation algorithm to generate TDI camera control commands.

[0008] The wafer surface is scanned according to the TDI camera control instructions and illumination mode instructions. Raw image data is collected and metadata labels are added. The data is input into the depth defect classification model and the defect information map and corresponding information are output.

[0009] The aforementioned detection method for imageless dark field detection in semiconductors integrates an adjustable laser source array, a high-performance TDI camera, and a pre-scanning sensor. The pre-scanning sensor acquires brightness information of the pre-scanning area and generates a pre-scanning reflectance map in real time. The method includes:

[0010] It integrates a pre-scan sensor and rigidly integrates it with the high-performance main TDI camera and adjustable laser light source array, and sets up an acquisition mechanism that is synchronized with the main TDI camera's row trigger signal;

[0011] Using a reference illumination mode, raw light intensity signals are acquired through a pre-scanning sensor and converted into standardized relative reflectance to generate a pre-scanning reflectance spectrum.

[0012] The aforementioned detection method for imageless dark field detection in semiconductors includes illuminating a pre-scanning area sequentially with illumination light of different wavelengths and incident angles using an adjustable laser light source array, generating multiple scan images, and extracting features to generate a multimodal differential feature field, comprising:

[0013] The lighting instruction sequence and sensor triggering sequence are loaded according to the preset detection formula. The adjustable laser source array is triggered by global hardware synchronization pulse to quickly switch the lighting mode according to the sequence and simultaneously acquire the corresponding scan image.

[0014] Feature operations are performed pixel-by-pixel in real time on multiple spatially aligned calibration images to generate a multimodal differential feature field.

[0015] The aforementioned detection method for semiconductor dark field detection without image coverage includes: identifying defects in a multimodal differential feature field using a multimodal illumination and defect feature co-enhancement algorithm; predicting the main defect types within the region; and generating corresponding illumination mode instructions based on the defect types. The method includes:

[0016] Based on the multimodal differential feature field, by evaluating its discriminative ability to candidate defects, iteratively generating and driving a discriminative illumination mode supplementary scan to expand the feature vector until the confidence level is met, and outputting the defect diagnosis tuple;

[0017] Based on the defect diagnosis tuple, the corresponding lighting mode is obtained through the preset defect-lighting response knowledge base. The final lighting mode is determined based on the confidence threshold, and its parameters are encapsulated to obtain the lighting mode instruction.

[0018] The aforementioned detection method for imageless dark field detection in semiconductors includes, based on a pre-scanned reflectance spectrum, calculating optimal TDI parameters using a TDI parameter dynamic modulation algorithm, and generating TDI camera control commands, comprising:

[0019] Based on the pre-scan reflectance spectrum, the extreme value of the integral number is calculated, and the optimal integral number is obtained by combining the strategy weights. The control parameter triplet is then generated based on this.

[0020] The parameter triplet is encoded into a specific camera register read / write instruction packet, and a cyclic redundancy check code is embedded to obtain the TDI camera control instruction.

[0021] The aforementioned detection method for imageless dark-field inspection of semiconductors includes scanning the wafer surface according to TDI camera control commands and illumination mode commands, acquiring raw image data and labeling it with metadata tags, inputting a depth defect classification model, and outputting a defect information map and corresponding information, including:

[0022] The wafer is scanned according to the bright mode and TDI camera control instructions, and raw image data and metadata containing spatial identifiers and acquisition parameters are generated in parallel.

[0023] By using a dual-branch deep neural network to extract the visual features of the original image and the physical context features encoded by the metadata, the feature weights are adjusted with the help of a gating mechanism to output a defect information map and corresponding information.

[0024] A high-speed camera for imageless dark-field detection in semiconductors, comprising:

[0025] The pre-scan spectrum generation module integrates an adjustable laser source array, a high-performance TDI camera, and a pre-scan sensor. It collects brightness information of the pre-scan area through the pre-scan sensor and generates a pre-scan reflectance spectrum in real time.

[0026] The multimodal feature field construction module is used to generate multiple scan images and extract features to generate a multimodal differential feature field by illuminating the pre-scanned area with illumination light of different wavelengths and incident angles through an adjustable laser light source array.

[0027] The lighting instruction generation module is used to identify defects in the multimodal differential feature field through a multimodal lighting and defect feature collaborative enhancement algorithm, predict the main defect types in the area, and generate corresponding lighting mode instructions based on the defect types.

[0028] The TDI control command generation module is used to calculate the optimal TDI parameters and generate TDI camera control commands based on the pre-scan reflectance spectrum and the TDI parameter dynamic modulation algorithm.

[0029] The defect classification module is used to scan the wafer surface according to the TDI camera control instructions and illumination mode instructions, collect raw image data and label metadata, input the deep defect classification model, and output defect information map and corresponding information.

[0030] The beneficial effects achieved by this invention are as follows:

[0031] This invention provides a method that can intelligently and dynamically adjust detection parameters, thereby significantly improving the signal-to-noise ratio and classification accuracy of defect detection while maintaining high-speed scanning. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0033] Figure 1 This is a flowchart of a detection method for imageless dark field detection of semiconductors provided in Embodiment 1 of this application.

[0034] Figure 2 This is a schematic diagram of a high-speed camera for semiconductor dark field detection provided in Embodiment 2 of this application. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Example 1

[0037] like Figure 1 As shown, Embodiment 1 of this application provides a detection method for patternless dark field detection in semiconductors, comprising:

[0038] S1: Integrates an adjustable laser light source array, a high-performance TDI camera, and a pre-scanning sensor. The pre-scanning sensor collects brightness information of the pre-scanning area and generates a pre-scanning reflectance map in real time.

[0039] The process integrates an adjustable laser source array, a high-performance TDI camera, and a pre-scanning sensor. The pre-scanning sensor collects brightness information from the pre-scanning area and generates a pre-scanning reflectance map in real time. This includes the following sub-steps:

[0040] S11: Integrates a pre-scan sensor and rigidly integrates it with the high-performance main TDI camera and adjustable laser light source array, and sets up an acquisition mechanism that is synchronized with the main TDI camera's row trigger signal;

[0041] Specifically, the pre-scan sensor is a high-speed, high-sensitivity linear optical sensor (such as a linear CMOS sensor) with a resolution typically lower than that of the main TDI camera. The pre-scan sensor is not a separate module but is rigidly integrated into the same scanning head as the main TDI camera and the adjustable laser light source array. Its optical field of view precisely leads the main TDI camera's field of view by a fixed distance (e.g., D millimeters) in the scanning motion direction (e.g., the X-axis). This arrangement ensures that it can always "see" the area the main camera is about to scan in advance. The system uses the line trigger signal of the main TDI camera as a unified timing reference. The acquisition frequency of the pre-scan sensor maintains a fixed ratio (e.g., N:1) with the main camera, and hardware circuitry ensures strict synchronization for each trigger. That is, for every line the main camera prepares to scan, the pre-scan sensor has already completed the acquisition and processing of N lines of the corresponding look-ahead area.

[0042] S12: Using the reference illumination mode, the raw light intensity signal is acquired by the pre-scanning sensor and converted into standardized relative reflectance to generate a pre-scanning reflectance spectrum.

[0043] During the pre-scanning phase, the tunable laser light source array is fixed to a specific "reference illumination mode." This mode is carefully selected (e.g., using a specific wavelength and moderate incident angle that are sensitive to most thin-film materials and not easily saturated) and its sole purpose is to stably and consistently excite the reflectivity of the wafer surface, not for defect imaging. The pre-scanning sensor then acquires the raw light intensity signal under this stable illumination. Subsequently, pre-stored system response calibration curves (including the light source intensity-sensor response curve, optical system attenuation coefficient, etc.) are invoked in real time to calculate and convert the raw grayscale values ​​into standardized relative reflectivity values ​​directly related to the physical properties of the surface material. These continuous data streams are stitched together in real time to directly generate a two-dimensional digital reflectivity matrix (i.e., a pre-scan reflectivity map), where each unit represents the calibrated reflectivity of a pixel.

[0044] S2: By using an adjustable laser light source array, the pre-scanning area is illuminated sequentially with illumination light of different wavelengths and incident angles to generate multiple scan images and extract features to generate a multimodal differential feature field;

[0045] The process involves using an adjustable laser light source array to sequentially illuminate the pre-scanning area with illumination light of different wavelengths and incident angles, generating multiple scanned images and extracting features to generate a multimodal differential feature field. This includes the following sub-steps:

[0046] S21: Load the lighting instruction sequence and sensor trigger timing according to the preset detection formula, and trigger the adjustable laser source array to quickly switch the lighting mode according to the sequence and simultaneously acquire the corresponding scan image through the global hardware synchronous pulse trigger;

[0047] Before the pre-scan begins, a complete illumination command sequence (e.g., `[Mode A (450nm, 20°), Mode B (650nm, 70°)]`) and the corresponding sensor trigger timing are pre-loaded into the local buffers of the light source array and the pre-scan sensor according to the preset detection formula. When the pre-scan sensor moves with the platform to the starting position of the target area, it emits a global hardware synchronization pulse. Under this pulse: the adjustable laser light source array immediately switches to Mode A and reaches stable output within microseconds. Almost simultaneously, the pre-scan sensor completes the acquisition of the first scan image under stable illumination in Mode A. Immediately afterwards, in the next precise clock cycle, the light source instantly switches to Mode B and stabilizes, and the sensor synchronously acquires the second image, and so on. In addition, at the instant of each image acquisition, the precise encoder position of the current platform is synchronously recorded, and this position information is directly bound to the image data row as metadata, ensuring that multiple images have sub-pixel level alignment accuracy in spatial coordinates.

[0048] S22: Perform feature calculations pixel-by-pixel on multiple spatially aligned calibration images in real time to generate a multimodal differential feature field.

[0049] The raw analog signals acquired by the pre-scanning sensor are immediately preprocessed in real time after analog-to-digital conversion. This includes independent background dark noise subtraction and non-uniformity gain correction for each illumination mode, and quantizing the results into standardized digital values. Then, pixel-by-pixel real-time differencing or other feature operations (such as ratio calculation) are performed on two (or more) calibrated, spatially aligned image data. For example, the pixel values ​​of the low-angle illumination image are subtracted from the corresponding pixel values ​​of the high-angle illumination image. This generates a multimodal difference feature field. Each "pixel" in this feature field no longer represents simple light intensity, but rather a feature vector. This vector describes the unique scattering behavior pattern of the surface point under various lighting conditions.

[0050] S3: For the multimodal differential feature field, the multimodal illumination and defect feature collaborative enhancement algorithm is used to identify defects, predict the main defect types in the area, and generate corresponding illumination mode instructions based on the defect types;

[0051] Specifically, for the multimodal differential feature field, defects are identified through a multimodal illumination and defect feature collaborative enhancement algorithm, the main defect types in the region are predicted, and corresponding illumination mode instructions are generated based on the defect types. This includes the following sub-steps:

[0052] S31: Based on the multimodal differential feature field, by evaluating its discriminative ability to candidate defects, iteratively generate and drive a discriminative illumination mode supplementary scan to expand the feature vector until the confidence level is met, and output the defect diagnosis tuple.

[0053] Based on the initial perception results, new and more discriminative illumination patterns are generated in real time to drive supplementary scanning. Through an iterative spectral decoupling process, defect types are accurately separated and identified. Specifically, a multimodal differential feature field is used as the core input, where each "pixel" of this feature field is a feature vector, for example... ,in The values ​​represent the normalized response values ​​of that point under predefined lighting modes A and B. This is its difference value. This vector describes the unique scattering behavior pattern of surface points under various known illuminations.

[0054] First, consider the current feature vector set. The analysis evaluates the discriminative power of the current feature vector for different candidate defect types (e.g., particles, scratches, stains). Specifically, the current feature vector is assigned to a lightweight classifier (e.g., a support vector machine or a small neural network), which outputs a probability distribution. For example: `P(particles) = 0.45, P(scratches) = 0.43, P(stains) = 0.12`. Next, the difference between the top-two probabilities of this distribution is calculated. If the difference is less than a threshold, it is determined that "the current lighting pattern combination provides insufficient information to reliably distinguish defects," requiring the introduction of additional lighting with specific spectral or angular attributes to acquire new discriminative features and resolve the confusion. Then, the parameters of the next lighting pattern that best reduces classification uncertainty are calculated. The specific formula is as follows: , This represents the parameters of the optimal discrimination lighting mode determined by the system after calculation; where, Represents the center wavelength of the laser source; Represents the angle of incidence of the laser beam; This indicates that the system should be implemented at all programmable wavelengths. and angle of incidence In the combinatorial space, a search and evaluation are performed, with the objective being... Maximize a specific set of parameters ; This means that for any given set of lighting conditions The system will iterate through all different candidate defect types. And identify the pair of defect types that have the most similar theoretical responses and are the most difficult to distinguish; : represents the square of the Euclidean distance; Indicates the defect type Under specific lighting conditions Below, the normalized theoretical scattering response vector of the defect is predicted; this vector may contain multi-dimensional features such as predicted intensity and polarization state changes; The difference in theoretical response vectors between the two types of defects under the same lighting conditions was calculated; Calculate the difference vector Norm squared, which is the sum of squares of the differences in each dimension, represents the separability of two types of defects in the theoretical feature space under specific illumination. The larger the value, the easier it is for the system to predict that the two are distinguishable; the smaller the value, the more similar the predicted features are, and the easier it is to confuse them.

[0055] Next, the adjustable laser source array is controlled to switch to the newly generated one within microseconds. The pattern is then activated, and the pre-scan sensor performs an extremely rapid supplemental scan of the area to obtain new response values. Then As a new dimension, it is added to the feature vector. In the next step, update the classifier and re-evaluate the confidence level. This loop can be iterated quickly until the confidence level meets the requirements or the time budget limit is reached.

[0056] After the iteration terminates, the system reclassifies and outputs the final determined main defect type based on the final expanded high-dimensional feature vector. And the quantized confidence level calculated based on the entropy of the classification probability distribution. At the same time, based on the identified defect type The theoretical optical model was used to calculate the theoretical contrast factor under the upcoming main illumination mode. The final output is a defect diagnosis tuple. .

[0057] S32: Based on the defect diagnosis tuple, obtain the corresponding lighting mode through the preset defect-lighting response knowledge base, determine the final lighting mode based on the confidence threshold, and encapsulate its parameters to obtain the lighting mode instruction.

[0058] Using the defect diagnosis tuple as input, a pre-defined defect-lighting response knowledge base is queried. For simple defects, a single optimal lighting mode may be directly mapped. For complex or composite defects, the mapping is a temporal composite illumination pattern sequence. This ensures that within the same integration time of the main scan, the response of the defect under multiple optimal conditions is obtained through rapid switching of the light source. Then, a flexible decision-making mechanism is used to synthesize the final instruction.

[0059] If confidence level (The highest threshold) then employs the optimal (or composite) lighting pattern mapped from the knowledge base. If the confidence level... If the minimum threshold is reached, the uncertain diagnostic result is abandoned, and a preset, broad-spectrum, robust lighting mode with moderate response to various defects is adopted instead. This ensures a basic detection rate. Finally, the determined illumination mode parameters (which may be single or sequential) are encapsulated to obtain the corresponding illumination mode instructions, and timestamped with the TDI control instructions to ensure that the illumination conditions are ready at the precise moment when the TDI camera scans the area.

[0060] S4: Calculate the optimal TDI parameters based on the pre-scan reflectance spectrum using the TDI parameter dynamic modulation algorithm, and generate TDI camera control commands;

[0061] S41: Calculate the extreme value of the integral number based on the pre-scan reflectance spectrum, combine it with the strategy weights to obtain the optimal integral number, and use it to generate the control parameter triplet;

[0062] Receive pre-scan reflectance map .in, It is a standardized relative reflectance value. This corresponds to the physical coordinates on the wafer. A TDI parameter dynamic modulation algorithm is executed to generate dynamic TDI parameters that strictly correspond one-to-one with each pixel in the pre-scanned reflectivity map. Specifically, this is achieved using a high-resolution, standardized pre-scanned reflectivity map. As input, each independent The value represents the reflection characteristics of a micro-region on the wafer surface and is embedded with the camera saturation threshold. System baseline noise Minimum signal-to-noise ratio requirement Prior knowledge, etc. Then for each... The values ​​are validated and filtered, and then two key calculation channels are executed in parallel and independently for each pixel:

[0063] The first channel aims to prevent signal saturation, based on the formula: ,in, Indicates the current scan point The maximum number of integration stages allowed to prevent signal saturation from the main TDI camera; This indicates the full-well capacity of the main TDI camera cell, which is the maximum number of photogenerated charges that a single cell can hold before saturation. This indicates background signal bias, which is caused by the dark current of the TDI camera itself and stray light that cannot be completely eliminated within the optical system. Indicates the wavelength of illumination The relevant system photoelectric response coefficient is quantified as the efficiency of generating charge on a camera pixel by a unit reflectivity signal under illumination of a specific wavelength. Indicates at wavelength Under illumination, coordinates The standardized reflectance value at that location is derived from the multispectral reflectance map during the pre-scanning phase; This represents the optical crosstalk correction term, used to quantify the impact of neighboring pixel signals on the current point due to diffraction and aberrations in the optical system. The "bleeding" effect; Represents local reflectivity distribution The two-dimensional Fourier transform converts it from the spatial domain to the spatial frequency domain. ; The modulation transfer function represents the system's modulation transfer function, which describes the degree of attenuation of signals at different spatial frequencies. This indicates that performing an inverse Fourier transform on the frequency signal after MTF attenuation yields the result of optical crosstalk at a point in the spatial domain. The additional signal increment generated at the location; It represents the weighting index for different spatial frequency components (microscopic spatial frequency information of the pre-scanned spectrum).

[0064] The second channel is designed to meet the minimum defect detection sensitivity, according to the formula: ,in, Indicates the current point , using specific When parameters are equal, the system's estimated signal-to-noise ratio for the target area is the core indicator for evaluating detection sensitivity; Indicates the effective signal strength; This represents the fundamental signal determined by the surface reflectivity and the integral series; Represents the normalized defect contrast factor For example, for high-contrast particles, For minor, shallow scratches, This factor directly injects the predicted information of the defect type into the signal-to-noise ratio calculation; This represents the square root of the total noise, which includes four main noise sources: The variance of photon shot noise is represented by a Poisson statistic, where... The noise factor of the sensor; The variance of the readout noise is generated by the camera readout circuit and is a fixed parameter that is independent of the signal magnitude. Represents the variance of fixed-mode noise; It is a two-dimensional map obtained through calibration, that is, by obtaining multiple frames of images through dark field calibration and averaging them, what is obtained is a fixed pattern noise "map" inherent to each pixel that does not change with the integration time, recording the inherent response non-uniformity of each pixel. Represents the variance of the quantization noise; This represents the minimum signal-to-noise ratio threshold for dynamic processing; its input... This is the confidence level for predicting the defect type in the current area; when the confidence level is high, the threshold can be appropriately relaxed to pursue speed; when the confidence level is low, the threshold should be tightened to ensure the detection rate. The minimum required integral series is then calculated. .

[0065] Subsequently within the feasible range Internally, based on a configurable policy weight (Preferential to speed or sensitivity) Arbitrate the optimal integral level specific to that pixel. .

[0066] Among them, strategy weights Before the system is put into use, a batch of known "gold standard" wafers representing typical processes are used to conduct full-parameter scanning tests. This is achieved by analyzing different reflectivity regions under different conditions. The actual defect detection rate and signal-to-noise ratio data under the combination were used to plot the Pareto front curve of the "speed-sensitivity" under this process.

[0067] Based on this, strictly adhere to the preset platform scanning speed. With pixel size The physical relationship is used to synchronously lock the line frequency corresponding to that point. And output the appropriate analog gain by looking up a table. Thus for each coordinate Generate a unique set of optimal control parameter triples. .

[0068] in, The integration series is the most crucial parameter in TDI cameras, representing the number of photosensitive units the charge passes through sequentially within the camera to accumulate the signal. The integration series directly and linearly determines the signal-to-noise ratio and sensitivity. (Line frequency...) This refers to the number of lines of image data output per second by the TDI camera. It must be strictly synchronized with the wafer platform's movement speed. Analog gain. This is the amplification factor applied during the analog circuit stage before the signal undergoes analog-to-digital conversion. It amplifies both the signal and noise simultaneously and is primarily used for fine-tuning. The signal-to-noise ratio is primarily changed by adjusting the integral stage; gain is a secondary means. A triplet is a direct driver. The core settings of the camera hardware.

[0069] S42: Encode the parameter triplet into a specific camera register read / write instruction packet and embed a cyclic redundancy check code to obtain the TDI camera control instruction.

[0070] Numerical parameters are transformed into reliable, executable hardware instructions strictly aligned with spatiotemporal coordinates. First, the parameter triplets are encoded into specific camera register read / write instruction packets according to industry-standard protocols (such as GenICam), and a cyclic redundancy check (CRC) code is embedded to ensure integrity. To achieve precise timing control, the system uses a fixed spatial lead distance between the pre-scan sensor and the main TDI camera. Combined with platform speed Accurately calculate the time offset by which the instruction needs to be issued in advance. (in (Inherent system latency). All command packets are placed into a hardware-scheduled forward command queue based on this timestamp. When the main scan position encoder triggers the execution condition, the corresponding command packet is sent to the TDI camera interface in real time with low latency. In addition, the system queries the camera status periodically via heartbeat packets to achieve closed-loop monitoring; if command execution is abnormal, it will automatically switch to the backup parameter chain and issue an alarm. Finally, a TDI camera control command stream that is strictly synchronized with the main scan motion and has fault tolerance is output.

[0071] S5: Scans the wafer surface according to the TDI camera control instructions and illumination mode instructions, acquires raw image data and labels metadata, inputs the depth defect classification model, and outputs defect information map and corresponding information.

[0072] The process involves scanning the wafer surface according to TDI camera control commands and illumination mode commands, acquiring raw image data and labeling it with metadata tags, inputting it into a depth defect classification model, and outputting a defect information map and corresponding information, including the following sub-steps:

[0073] S51: Scans the wafer according to the bright mode and TDI camera control instructions, and generates raw image data and metadata containing spatial identifiers and acquisition parameters in parallel;

[0074] The system strictly follows the illumination mode instructions and TDI camera control instructions to control the hardware execution. At time t, when the scanning platform moves to the wafer coordinates... The adjustable light source array has been switched to a specific mode optimized for predicting defect type T, as instructed. Simultaneously, the main TDI camera has been dynamically adjusted according to instructions to match the reflectivity of that point. Matching integral series Line frequency and gain G.

[0075] After scanning begins, while the main TDI camera generates the raw image data stream, the system simultaneously generates a fully synchronized metadata stream. This metadata annotates each data unit (such as an image patch or a row of pixels) with a structured tag package, containing at least: spatial identifier (absolute wafer coordinates). ), Parameters used when performing this data collection Then, the system packages and encapsulates the raw image data with the corresponding metadata tag package to form a self-describing raw image data block containing context information.

[0076] S52: The original image visual features and metadata encoded physical context features are extracted by a dual-branch deep neural network. The feature weights are adjusted by a gating mechanism, and the defect information map and corresponding information are output.

[0077] The network processes raw image data blocks containing contextual information using a specially designed deep neural network. Specifically, the network employs a two- or multi-branch structure, with the image feature extraction branch processing the raw image through a convolutional neural network to extract visual features. The context feature fusion branch uses metadata... After the parameters are encoded, they are input into a fully connected network to generate physical context features. The core lies in the existence of an adaptive feature gating mechanism within the network, which dynamically adjusts the feature gating based on the predicted defect type and confidence level. and The weight in the final classification decision. For example, when At lower levels, the model relies more on the original image features. This reduces the impact of potentially erroneous predictive context; when When the defect type is high and sensitive to light, the improvement is significant. Medium lighting mode The weights of the features are then assigned. Based on this, the two sets of features are weighted and fused to generate joint features. The joint features are then processed by the model's intrinsic classifier (fully connected layers and a softmax function), outputting the final defect category and a confidence score calibrated with context information. Each confirmed defect node is associated with a complete data package, including: final classification, coordinates, size, image slice, and crucial process traceability information—namely, the precise acquisition parameters that produced the clear image of the defect and upstream prediction records. For example, if a confirmed particle is found at coordinates (X,Y), the image is taken under blue low-angle light optimized for 'particles'. Below, high signal-to-noise ratio parameters are used. The data was collected and is consistent with the pre-scan prediction.

[0078] Example 2

[0079] like Figure 2 As shown, Embodiment 2 of this application provides a high-speed camera for semiconductor imageless dark field detection, comprising:

[0080] The pre-scanning spectrum generation module 21 integrates an adjustable laser light source array, a high-performance TDI camera, and a pre-scanning sensor. It collects brightness information of the pre-scanning area through the pre-scanning sensor and generates a pre-scanning reflectance spectrum in real time.

[0081] The multimodal feature field construction module 22 uses an adjustable laser light source array to illuminate the pre-scanning area with illumination light of different wavelengths and incident angles in succession, generating multiple scan images and extracting features to generate a multimodal differential feature field;

[0082] The lighting instruction generation module 23 identifies defects in the multimodal differential feature field through a multimodal lighting and defect feature collaborative enhancement algorithm, predicts the main defect types in the area, and generates corresponding lighting mode instructions based on the defect types.

[0083] The TDI control command generation module 24 calculates the optimal TDI parameters and generates TDI camera control commands based on the pre-scan reflectance spectrum and the TDI parameter dynamic modulation algorithm.

[0084] The defect classification module 25 scans the wafer surface according to the TDI camera control instructions and illumination mode instructions, collects raw image data and labels metadata, inputs the deep defect classification model, and outputs defect information map and corresponding information.

[0085] Corresponding to the above embodiments, the present invention provides a computer storage medium, including: at least one memory and at least one processor;

[0086] The memory is used to store one or more program instructions;

[0087] A processor is used to run one or more program instructions to execute a detection method for imageless dark-field detection of semiconductors.

[0088] Corresponding to the above embodiments, this embodiment of the invention provides a computer-readable storage medium containing one or more program instructions, which are executed by a processor to provide a detection method for semiconductor dark-field detection without image.

[0089] The embodiments disclosed in this invention provide a computer-readable storage medium storing computer program instructions that, when executed on a computer, cause the computer to perform the aforementioned detection method for semiconductor dark-field detection without patterns.

[0090] In this embodiment of the invention, the processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0091] The various methods, steps, and logic diagrams disclosed in the embodiments of this invention can be implemented or executed. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The processor reads information from the storage medium and, in conjunction with its hardware, completes the steps of the above methods.

[0092] The storage medium can be memory, such as volatile memory or non-volatile memory, or may include both volatile and non-volatile memory.

[0093] Among them, non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory.

[0094] Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (Synchlink DRAM, SLDRAM), and direct memory bus RAM (DRRAM).

[0095] The storage media described in the embodiments of the present invention are intended to include, but are not limited to, these and any other suitable types of memory.

[0096] Those skilled in the art will recognize that, in one or more of the examples above, the functions described in this invention can be implemented using a combination of hardware and software. When applied as software, the corresponding functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of computer programs from one place to another. Storage media can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0097] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.

Claims

1. A detection method for patternless dark-field detection in semiconductors, characterized in that, include: It integrates an adjustable laser light source array, a high-performance TDI camera, and a pre-scanning sensor. The pre-scanning sensor collects brightness information of the pre-scanning area and generates a pre-scanning reflectance map in real time. By using an adjustable laser light source array, the pre-scanning area is illuminated sequentially with illumination light of different wavelengths and incident angles to generate multiple scan images and extract features to generate a multimodal differential feature field. For multimodal differential feature fields, a multimodal illumination and defect feature collaborative enhancement algorithm is used to identify defects, predict the main defect types in the region, and generate corresponding illumination mode instructions based on the defect types. Based on the pre-scan reflectance spectrum, the optimal TDI parameters are calculated using a TDI parameter dynamic modulation algorithm to generate TDI camera control commands. Scan the wafer surface according to the TDI camera control command and illumination mode command, collect raw image data and label metadata, input the deep defect classification model, and output defect information map and corresponding information; Based on the pre-scan reflectivity map, the optimal TDI parameters are calculated using a TDI parameter dynamic modulation algorithm, and TDI camera control commands are generated, including: Based on the pre-scan reflectance spectrum, the extreme value of the integral number is calculated, and the optimal integral number is obtained by combining the strategy weights. The control parameter triplet is then generated accordingly. The parameter triplet is encoded into a specific camera register read / write instruction packet, and a cyclic redundancy check code is embedded to obtain the TDI camera control instruction.

2. The detection method for patternless dark field detection of semiconductors according to claim 1, characterized in that, It integrates an adjustable laser source array, a high-performance TDI camera, and a pre-scanning sensor. The pre-scanning sensor collects brightness information from the pre-scanning area and generates a pre-scanning reflectance map in real time, including: It integrates a pre-scan sensor and rigidly integrates it with the high-performance main TDI camera and adjustable laser light source array, and sets up an acquisition mechanism that is synchronized with the main TDI camera's row trigger signal; Using a reference illumination mode, raw light intensity signals are acquired through a pre-scanning sensor and converted into standardized relative reflectance to generate a pre-scanning reflectance spectrum.

3. The detection method for imageless dark field detection of semiconductors according to claim 1, characterized in that, By using an adjustable laser light source array, the pre-scanned area is illuminated sequentially with illumination light of different wavelengths and incident angles to generate multiple scanned images and extract features to generate a multimodal differential feature field, including: The lighting instruction sequence and sensor triggering sequence are loaded according to the preset detection formula. The adjustable laser source array is triggered by global hardware synchronization pulse to quickly switch the lighting mode according to the sequence and simultaneously acquire the corresponding scan image. Feature operations are performed pixel-by-pixel in real time on multiple spatially aligned calibration images to generate a multimodal differential feature field.

4. The detection method for patternless dark field detection of semiconductors according to claim 1, characterized in that, For multimodal differential feature fields, a multimodal illumination and defect feature co-enhancement algorithm is used to identify defects, predict the main defect types in the region, and generate corresponding illumination mode instructions based on the defect types, including: Based on the multimodal differential feature field, by evaluating its discriminative ability to candidate defects, iteratively generating and driving a discriminative illumination mode supplementary scan to expand the feature vector until the confidence level is met, and outputting the defect diagnosis tuple; Based on the defect diagnosis tuple, the corresponding lighting mode is obtained through the preset defect-lighting response knowledge base. The final lighting mode is determined based on the confidence threshold, and its parameters are encapsulated to obtain the lighting mode instruction.

5. The detection method for patternless dark field detection of semiconductors according to claim 1, characterized in that, The wafer surface is scanned according to TDI camera control commands and illumination mode commands. Raw image data is acquired and metadata labels are added. The data is then input into a depth defect classification model, and the defect information map and corresponding information are output, including: The wafer is scanned according to the bright mode and TDI camera control instructions, and raw image data and metadata containing spatial identifiers and acquisition parameters are generated in parallel. By using a dual-branch deep neural network to extract the visual features of the original image and the physical context features encoded by the metadata, the feature weights are adjusted with the help of a gating mechanism to output a defect information map and corresponding information.

6. A high-speed camera for imageless dark-field detection in semiconductors, characterized in that, include: The pre-scan spectrum generation module integrates an adjustable laser source array, a high-performance TDI camera, and a pre-scan sensor. It collects brightness information of the pre-scan area through the pre-scan sensor and generates a pre-scan reflectance spectrum in real time. The multimodal feature field construction module is used to generate multiple scan images and extract features to generate a multimodal differential feature field by illuminating the pre-scanned area with illumination light of different wavelengths and incident angles through an adjustable laser light source array. The lighting instruction generation module is used to identify defects in the multimodal differential feature field through a multimodal lighting and defect feature collaborative enhancement algorithm, predict the main defect types in the area, and generate corresponding lighting mode instructions based on the defect types. The TDI control command generation module is used to calculate the optimal TDI parameters and generate TDI camera control commands based on the pre-scan reflectance spectrum and the TDI parameter dynamic modulation algorithm. The TDI control instruction generation module is specifically used to calculate the extreme value of the integral number based on the pre-scan reflectivity spectrum, combine it with the strategy weight to obtain the optimal integral number, and generate a control parameter triplet based on this; the parameter triplet is encoded into a specific camera register read / write instruction package, and a cyclic redundancy check code is embedded to obtain the TDI camera control instruction; The defect classification module is used to scan the wafer surface according to the TDI camera control instructions and illumination mode instructions, collect raw image data and label metadata, input the deep defect classification model, and output defect information map and corresponding information.

7. A computer-readable storage medium, characterized in that, It includes one or more program instructions, which are executed by a processor as described in any one of claims 1-5, for a detection method for semiconductor dark-field detection without image.

Images