An AI-based method for analyzing and classifying chip test failure modes

Abstract

This invention belongs to the field of industrial vision inspection technology and relates to an artificial intelligence-based method for analyzing and classifying chip testing failure modes. The method includes the following steps: training a multi-label classifier using occlusion enhancement based on historical failure maps and multi-label vectors to obtain a trained multi-label classifier; acquiring the wafer failure map to be classified and feeding it into the trained multi-label classifier for inference to obtain an initial probability vector; dividing the wafer failure map into a uniform grid to construct a failure density map and calculating the average density of the entire map; performing connected component analysis on the wafer failure map and calculating the average density of each failure connected component using the failure density map. This method effectively solves the problem of secondary failure modes being suppressed by dominant failures in mixed failure scenarios, significantly improving the recall rate of secondary failure modes, and automatically refining the decision threshold as production line data accumulates.

Description

Technical Field

[0001] This invention belongs to the field of industrial vision inspection technology and relates to an artificial intelligence-based method for analyzing and classifying chip testing failure modes. It is used for failure map analysis in semiconductor wafer mass production testing, solves the problem of missed detection of mixed failures, improves the recall rate of minor failures, and refines the threshold. Background Technology

[0002] In the mass production of semiconductor wafers, after wafer-level electrical testing, ATE (Automatic Test Equipment) equipment generates a failure distribution map for each wafer, recording the test pass and failure status of each chip. The spatial pattern of the failure distribution map highly corresponds to the specific process defect type, serving as the core basis for process engineers to conduct root cause analysis and quickly locate process anomalies. As production line capacity increases and the number of wafers per batch continues to rise, higher demands are placed on the efficiency and accuracy of failure mode classification, especially in mixed failure scenarios where multiple failure modes coexist on the same failure distribution map, significantly increasing the difficulty of classification.

[0003] Currently, the industry-standard automated failure mode classification (FMM) scheme is based on deep convolutional neural networks to build single-label or multi-label classifiers. It uses a large amount of manually labeled historical failure distribution maps as training data, and through supervised learning, the classifier learns the spatial characteristics of various failure modes. During the inference phase, a forward inference is performed on the failure distribution map of a new wafer. Based on the comparison between the output probability values ​​and fixed decision thresholds, the final failure mode classification label is output. This scheme achieves high classification accuracy in single failure mode scenarios, has a mature deployment process, and has been applied in multiple mass production lines.

[0004] However, in mixed failure scenarios, large-area, dense failure modes have a significant dominance over the overall image features during convolutional feature extraction. This causes the feature responses of minor failure modes with smaller area proportions to be suppressed by the dominant failure modes, resulting in their output probabilities generally falling below the fixed decision threshold, thus causing systematic missed detections. Existing solutions cannot distinguish between the absence of minor failure modes and the lower probability of minor failure modes due to suppression, which are fundamentally different situations. Furthermore, they cannot adaptively adjust the decision threshold based on the differences in suppression depth between different dominant and minor failure categories. This leads to a persistently high missed detection rate for minor failure modes in mixed failure scenarios, severely impacting the directionality and accuracy of root cause analysis.

[0005] To address the industry demand for wafer failure mode classification, relevant technological research has been conducted, resulting in several patented solutions. Among them, Chinese patent CN113627457B discloses a method and system for classifying wafer defects using wafer defect images based on deep learning. This method employs multiple imaging modes to acquire wafer defect images, strings multiple machine learning models together using a directed acyclic graph architecture, optimizes model training by combining reference images, and completes defect classification based on the model's output path. However, this solution does not address the feature suppression problem of dominant defects on secondary defects in mixed failure scenarios, fails to solve the systematic under-detection problem of secondary failure modes, lacks an adaptive decision threshold, and has insufficient robustness in identifying mixed defect combinations.

[0006] Chinese patent application CN115984235A discloses a method and system for identifying mixed defects in wafer images based on image segmentation. It extracts single-defect regions through connected component analysis to create pixel-level labels, mixes them to generate a multi-defect dataset, and trains a U-Net semantic segmentation model to achieve end-to-end segmentation and recognition of mixed defects in wafer images. However, this scheme relies on pixel-level segmentation labels, resulting in extremely high annotation costs. Furthermore, it fails to decouple the feature suppression effect of dominant defects, making it impossible to recover weaker features of secondary defects through quantification of the suppression effect. This also leads to insufficient recognition accuracy in mixed scenes with significantly varying defect proportions.

[0007] Chinese patent CN120431093B discloses a mask-based wafer defect classification system and method. It concatenates a wafer image with defect mask channels to form a four-channel input, adapts it to a semantic segmentation model via a 1×1 convolutional layer, and completes defect segmentation and classification using a U-Net model. It also sets up dual inference paths to adapt to different application scenarios. However, this solution only optimizes segmentation accuracy through mask information fusion and does not design a dominant defect occlusion and secondary inference mechanism for mixed failure scenarios. It cannot quantify and recover suppressed secondary defect signals, and still suffers from the industry pain point of missing secondary defects in mixed defect scenarios.

[0008] In summary, existing solutions have failed to fundamentally solve the problem of systematic omissions caused by the suppression of secondary failure modes by the dominant failure mode features in mixed failure scenarios. They cannot balance classification efficiency with recognition accuracy in mixed failure scenarios, and are unable to meet the high reliability and high accuracy application requirements of wafer failure mode classification in semiconductor mass production lines. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the prior art, solve the technical problem that the secondary failure mode is suppressed by the dominant failure in the above-mentioned mixed failure scenario and is systematically missed, and provide an artificial intelligence-based method for analyzing and classifying chip test failure modes.

[0010] To achieve the above-mentioned objectives, this invention provides a method for analyzing and classifying chip testing failure modes based on artificial intelligence, comprising the following steps: Based on historical failure maps and multi-label vectors, a multi-label classifier is trained using occlusion enhancement to obtain a trained multi-label classifier. Obtain the wafer failure map to be classified, feed it into the trained multi-label classifier for inference, and obtain the initial probability vector; divide the wafer failure map into uniform grids, construct the failure density map, and calculate the average density of the entire map; Perform connected component analysis on the wafer failure map, calculate the average density of each failure connected component using the failure density map, set the pixels of the failure connected components with an average density higher than the average density of the entire image to zero, and obtain the occlusion image; feed the occlusion image into a trained multi-label classifier for inference to obtain the occlusion probability vector; calculate the difference between the occlusion probability vector and the initial probability vector to obtain the suppression recovery amount; A suppression relation matrix is ​​constructed based on historical inference data. The suppression relation matrix is ​​multiplied by a preset safety coefficient to obtain an adaptive threshold. When the probability of a certain class in the initial probability vector is not lower than the preset probability threshold, or the corresponding suppression recovery exceeds the adaptive threshold, the final label of that class is set to 1, and the final multi-label classification result is output.

[0011] This invention can scientifically quantify the masked secondary failure signals by intelligently masking high-density failure areas and performing secondary inference. Combined with adaptive thresholds driven by historical data, it can make dual-path decisions, which greatly improves the recall rate of weak feature defects in mixed failure modes and provides process engineers with a more comprehensive and accurate basis for process anomaly analysis.

[0012] The multi-label classifier training method using occlusion enhancement described in this invention includes: in each training batch, randomly selecting a portion of training samples with a preset enhancement probability; performing connected component analysis on the selected training samples, randomly selecting one of the failed connected components and setting its pixels to zero to obtain enhanced samples, the labels of the enhanced samples being consistent with the original training samples; and inputting the enhanced samples together with the unselected training samples in the current batch to update the parameters of the multi-label classifier.

[0013] This invention introduces a random occlusion enhancement mechanism for failed connected components during the model training phase. Compared with common image augmentation methods, this approach is more in line with the physical logic of chip failure analysis, enabling the classifier to learn in advance how to extract global correlation features when some features are lost. This ensures that the model has extremely high robustness and consistency when processing artificially constructed occluded images during the inference phase, thereby guaranteeing the authenticity of subsequent suppression recovery calculations.

[0014] The construction of the failure density map according to the present invention includes: dividing the wafer failure map into M×M spatial cells, wherein the value of M is such that the average number of chips in each spatial cell is not less than a preset minimum number of chips; calculating the ratio of the number of failed pixels in each spatial cell to the total number of pixels in that spatial cell to obtain the cell density value of each spatial cell; the cell density values ​​of all spatial cells constitute the failure density map; and averaging the cell density values ​​of all spatial cells to obtain the average density value of the entire map.

[0015] This invention employs a failure density map construction method based on a uniform grid. Compared to traditional coarse-grained full-map pixel statistics, this method can more precisely characterize the spatial clustering characteristics of failure points on the wafer. By quantifying the density of each cell and calculating the mean of the entire map, a standardized measurement of the degree of failure clustering can be achieved, providing an objective quantitative basis for subsequently accurately locating and removing the dominant dense failure regions.

[0016] The present invention describes a method where, when the average density of all failed connected components is not higher than the average density of the entire image, the pixels of the failed connected component with the highest average density are set to zero to obtain an occluded image.

[0017] This invention provides a degradation processing strategy for low yield or globally distributed failure scenarios. It ensures that even when the density of all failure regions is low, the system can still attempt to uncover potential secondary failures by stripping the largest failure connectivity. This fallback logic guarantees the continuity and stability of the analysis process under extreme abnormal conditions and avoids algorithm failure in complex production environments.

[0018] The construction of a suppression relationship matrix based on historical inference data as described in this invention includes: extracting a set of historical samples from historical inference data where the dominant failure category is the j-th type and the k-th type of failure actually exists, wherein the dominant failure category is the category with the highest probability in the initial probability vector of the corresponding historical sample; calculating the mean value of the suppression recovery amount corresponding to each historical sample in the historical sample set to obtain the element in the j-th row and k-th column of the suppression relationship matrix; the element in the j-th row and k-th column of the suppression relationship matrix is ​​positively correlated with the corresponding adaptive threshold, that is, the larger the element in the j-th row and k-th column, the larger the corresponding adaptive threshold.

[0019] This invention establishes a typical suppression model among different failure modes using historical production line reasoning data. By statistically analyzing the average suppression recovery under different failure category combinations, it can identify which dominant failures have a stronger interference with which secondary failures. This gives the decision threshold category awareness, making the decision process more consistent with the objective laws of physical failure and improving the scientific nature of the classification.

[0020] The method for constructing a suppression relationship matrix based on historical inference data as described in this invention further includes: counting the number of historical samples in the historical sample set; when the number is less than the preset minimum number of samples, setting the adaptive threshold of the j-th dominant failure to the k-th failure as the preset cold start threshold; when the number is not less than the preset minimum number of samples, switching the preset cold start threshold to the product of the element in the j-th row and k-th column of the suppression relationship matrix and the preset safety coefficient to obtain the corresponding adaptive threshold.

[0021] The preset safety factor described in this invention is determined in the following way: the recall rate of each failure mode in the validation set is not lower than the corresponding recall rate when a fixed threshold scheme is used as a constraint; within the range of values ​​that satisfy the constraint, the preset safety factor is obtained by searching with the goal of maximizing the average F1 score of the validation set.

[0022] After outputting the final multi-label classification result, the present invention further includes: recording the suppression recovery amount corresponding to the current wafer failure map and the final multi-label classification result into historical inference data; and re-statistically calculating the suppression relationship matrix based on the accumulated historical inference data according to a preset update cycle, and updating the adaptive threshold.

[0023] The multi-label classifier of this invention includes: a residual convolutional network for extracting residual features layer by layer from the input wafer failure map and outputting a feature vector after global average pooling; a fully connected layer for mapping the feature vector to a K-dimensional output vector, where K is the total number of predefined failure mode categories; and K independent activation functions for independently calculating each dimension of the K-dimensional output vector and outputting the existence probability of each failure mode. The existence probabilities of each failure mode are independent of each other and are not subject to the normalization constraints of other category probabilities.

[0024] The training loss used in training the multi-label classifier using the occlusion enhancement method described in this invention is determined as follows: For each of the K failure mode categories, the binary cross-entropy between the existence probability of each category and the corresponding multi-label vector component is calculated, and the average of the binary cross-entropy of all categories is taken to obtain the training loss; the parameters of the multi-label classifier are updated using the training loss as a supervision signal until the average F1 score on the validation set no longer improves.

[0025] Compared with existing technologies, this invention offers at least the following advantages: Addressing the issue of secondary failures being suppressed by dominant failures in mixed wafer failure scenarios, this invention enhances the multi-label classifier's reasoning ability on locally zeroed images through occlusion-enhanced training. Then, it locates the dominant failure's connected components using a failure density map and constructs occluded images for secondary reasoning, using the difference between the two reasoning results as the suppression recovery factor. Combining this with a suppression relationship matrix statistically obtained from historical reasoning data, it generates different adaptive thresholds for combinations of different dominant and secondary failure categories. Finally, it corrects the missed detection of secondary failures through dual-path decision-making. This significantly improves the recall rate of secondary failure modes in mixed failure scenarios, and the decision thresholds are automatically refined with the accumulation of production line data, continuously reducing the directional bias in root cause analysis. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the method for analyzing and classifying chip testing failure modes based on artificial intelligence, which relates to the present invention.

[0027] Figure 2 This is a schematic diagram comparing the accuracy of failure mode classification involved in this invention.

[0028] Figure 3 This is a heat map of the suppression signal recovery distribution involved in the present invention. Detailed Implementation

[0029] The technical solution of the present invention will now be clearly and completely described in conjunction with the embodiments and accompanying drawings.

[0030] The AI-based chip testing failure mode analysis and classification method provided in this embodiment is as follows: Figure 1 As shown, its main process includes the following steps: S101, occlusion-enhanced training of multi-label classifier.

[0031] Based on historical failure maps and multi-label vectors, a multi-label classifier is trained using occlusion enhancement to obtain a trained multi-label classifier.

[0032] This embodiment applies to scenarios where, after wafer-level electrical testing is completed during semiconductor wafer mass production, automatic failure mode analysis and classification of wafer failure maps are performed. Taking an 8-inch production line as an example, after the ATE equipment completes testing on a batch of wafers, it exports a large number of wafer failure maps, a significant proportion of which exhibit mixed failure characteristics, meaning that multiple failure modes, such as edge ring failures and center circular failures, exist simultaneously in the same wafer failure map. Process engineers need to complete the failure mode classification of all wafer failure maps within a specified time and report it to the process department. Manually interpreting each map is too time-consuming and has a high misclassification rate for mixed modes; therefore, a method that can automatically perform multi-label classification is needed.

[0033] First, historical wafer failure image data from the production line was collected to construct a training dataset. Each wafer failure image corresponds to a binary pixel image automatically generated by ATE equipment after all chips on the wafer have been tested. Pixels corresponding to failed chips are set to 1, and pixels corresponding to successful chips are set to 0. All wafer failure images were uniformly scaled to 224×224 pixels, and pixel values ​​were normalized to the range of 0 to 1. For each wafer failure image, process engineers labeled its corresponding multi-label vector. Satisfying the relation:

[0034] In the formula, The total number of predefined failure mode categories, This indicates that there is a first failure in this wafer failure diagram. Class failure modes, This indicates that the wafer failure map does not exist; multiple failure maps can exist simultaneously on the same wafer. This means that multiple tags can coexist. The preferred reference value is 8, which covers 8 failure modes: Edge-Ring, Center, Scratch, Random, Donut, Local, Linear, and Blob. The value is determined based on the actual failure mode types on the production line.

[0035] The difference between multi-label vectors and traditional single-label annotation is that if both Edge-Ring and Center failures exist on the same wafer failure map, the multi-label vector component corresponding to the Edge-Ring class is 1 and the multi-label vector component corresponding to the Center class is 1, rather than forcibly selecting a primary category. This is the data foundation for supporting mixed failure mode recognition.

[0036] After constructing the training dataset, a multi-label classifier network structure needs to be built to provide a foundational model for subsequent occlusion enhancement training. The multi-label classifier consists of a residual convolutional network, fully connected layers, and K independent activation functions. The residual convolutional network extracts residual features layer by layer from the input wafer failure map and outputs a feature vector after global average pooling. The fully connected layers map the feature vector to a K-dimensional output vector, where K is the total number of predefined failure mode categories. The K independent activation functions independently calculate the probability of each failure mode in the K-dimensional output vector, outputting the probability of each failure mode. The probability of each failure mode is independent of each other and is not subject to the normalization constraints of other category probabilities.

[0037] Specifically, ResNet-50 is used as the backbone of the residual convolutional network. A 224×224 pixel wafer failure map is input, and residual features are extracted layer by layer by the residual network. After global average pooling, a 2048-dimensional feature vector is output, which is then mapped to a K-dimensional output vector through a fully connected layer. Each dimension is activated by an independent Sigmoid activation function to output the probability of existence for that type of failure. Using an independent Sigmoid activation function instead of a Softmax activation function ensures that the probability outputs for each type of failure are independent and not subject to the normalization constraints of other class probabilities. This is a key structural design feature supporting multi-label output.

[0038] The training loss used to train the multi-label classifier is determined as follows: For each of the K failure mode categories, the binary cross-entropy between the existence probability of each category and its corresponding multi-label vector component is calculated. The mean of the binary cross-entropy for all categories is then taken to obtain the training loss. The standard binary cross-entropy loss function is used for independent supervision of the K labels, and the training loss satisfies the following relationship:

[0039] In the formula, This is the loss value. For the first The multi-label vector components corresponding to the class failure modes For the first The probability of the existence of a failure mode. The total number of failure mode categories is predefined. The parameters of the multi-label classifier are updated using the training loss as a supervision signal until the average F1 score on the validation set no longer improves. The optimizer is Adam, and the initial learning rate is set to... The optimal number of training rounds is 50, and the F1 score of each category is monitored on the validation set in each round.

[0040] Understandably, by calculating the binary cross-entropy independently for each of the K categories and then averaging the results, each failure mode participates equally in gradient updates. This avoids the category with more labels dominating the training direction in mixed failure scenarios, thus allowing the multi-label classifier to be fully trained to recognize the few coexisting categories.

[0041] To make the multi-label classifier more robust to locally zeroed images during the subsequent occlusion re-inference stage, occlusion enhancement is introduced during the training phase. Training the multi-label classifier with occlusion enhancement includes: randomly selecting a portion of training samples in each training batch with a preset enhancement probability; performing connected component analysis on the selected training samples, randomly selecting one invalid connected component and zeroing its pixels to obtain enhanced samples, whose labels remain consistent with the original training samples; and inputting the enhanced samples along with the unselected training samples from the current batch to update the parameters of the multi-label classifier.

[0042] It should be noted that the preferred reference value for the preset augmentation probability is 0.3, which can be adjusted within the range of 0.2 to 0.4 based on the F1 score on the validation set. In each training batch, a preset augmentation probability of 0.3 means that approximately 30% of the training samples will undergo occlusion augmentation before participating in the parameter update of this batch, while the remaining approximately 70% of the training samples will participate in training in their original state. Together, they constitute the complete input of this batch, ensuring a balanced distribution of occluded samples and original samples in the training batch. Keeping the labels of the augmented samples after zeroing consistent with the original training samples means that, physically, after the local failure connected region is occluded, the failure mode in the remaining areas of the wafer does not disappear. Therefore, the overall failure mode label should not change accordingly, and the multi-label classifier thus learns to determine the global failure mode category from the locally visible area.

[0043] Understandably, after the multi-label classifier has been exposed to wafer failure maps with locally connected components set to zero during the training phase, its response to artificially constructed occluded images during the inference phase is consistent with that of normal images. This avoids the occluded images being used as out-of-distribution samples, which would cause distortion in the inference results and thus ensure the reliability of the suppressed recovery amount in subsequent occluded re-inference.

[0044] Through the occlusion-enhanced training mechanism, the multi-label classifier can not only complete the initial inference of the complete wafer failure map, but also has a stable inference ability for subsequent occluded images. This provides a model-level guarantee for the comparability of the results of the two subsequent inferences, and fundamentally supports the core logic of recovering secondary failure signals based on differential operations.

[0045] S102, infer the initial probability, construct the density map and calculate the mean.

[0046] Obtain the wafer failure map to be classified, feed it into the trained multi-label classifier for inference, and obtain the initial probability vector; divide the wafer failure map into uniform grids, construct the failure density map, and calculate the average density of the entire map.

[0047] In this embodiment, after the multi-label classifier training converges, the wafer failure map to be classified is obtained from the production line, scaled to 224×224 pixels, and then fed into the trained multi-label classifier to perform the first forward inference, obtaining an initial probability vector. The initial probability vector contains K components, each component corresponding to the probability of the existence of a failure mode. The initial probability vector is denoted as... ,in For the first The probability of the existence of a failure mode, dimensionless, with a value ranging from 0 to 1.

[0048] While obtaining the initial probability vector, a failure density map needs to be constructed on the same wafer failure map to quantify the degree of failure clustering in various local regions of the wafer, providing a clear quantitative basis for subsequent identification of dominant failure connected regions. Constructing the failure density map includes: dividing the wafer failure map into M×M spatial cells, where M is chosen such that the average number of chips in each spatial cell is not less than a preset minimum number of chips; calculating the ratio of the number of failed pixels in each spatial cell to the total number of pixels in that spatial cell to obtain the cell density value for each spatial cell; the cell density values ​​of all spatial cells constitute the failure density map; and averaging the cell density values ​​of all spatial cells to obtain the average density of the entire map.

[0049] Specifically, the wafer failure map is divided into M×M spatial cells using a uniform grid. The preferred reference value for M is 14, corresponding to an average of no less than 4 chips in each spatial cell. This value can be adjusted based on the actual wafer size and chip size, with a preset minimum chip count of 4 being the preferred reference value. The cell density values ​​of cell number 1 satisfy the following relationship:

[0050] In the formula, For the first The cell density value of the specified space cell is dimensionless and ranges from 0 to 1. For the first The number of invalid pixels within the specified cell, in units of pixels; This represents the total number of pixels within the given cell, expressed in pixels. hour, indivual; The closer the value is to 1, the more densely packed the failed chips are in that local area. The closer to 0, the sparser the failure chips are in that local area.

[0051] Understandably, replacing the original pixel count with normalized density eliminates the difference in effective pixel counts between different spatial cells caused by edge clipping, making the degree of failure clustering in various local areas of the wafer directly comparable. For example, the spatial cell density values ​​corresponding to edge-ring failure regions are generally close to 1.0, while the spatial cell density values ​​corresponding to randomly scattered failure regions are generally below 0.2, and the difference between the two types of failures is clearly distinguishable on the failure density map.

[0052] The cell density values ​​of all M×M spatial cells constitute a failure density map, denoted as [missing information - likely a typology or symbol]. Simultaneously calculate the mean density of the entire map, which satisfies the following relationship:

[0053] In the formula, The density mean of the entire image is dimensionless. This represents the total number of cells in the space. For the first The cell density value of the specified spatial cell; the summation range covers all M×M spatial cells.

[0054] Understandably, the average density of the entire image reflects the overall failure clustering level of this wafer and serves as a benchmark reference value for subsequent determination of dominant failure connected regions: failure connected regions with an average density higher than the average density of the entire image have a failure clustering degree that exceeds the average level of the entire image and dominate in the feature extraction of the entire image, which is the fundamental reason why secondary failures are suppressed.

[0055] By performing two parallel operations on the wafer failure map—initial inference and failure density map construction—we simultaneously obtained the initial confidence distribution of various failure modes and the failure clustering distribution in the wafer spatial dimension, laying a data foundation for subsequent accurate identification of dominant failure connected components and implementation of occlusion operations.

[0056] S103, analyze the connected components and calculate the difference value to suppress the recovery amount.

[0057] Connectivity analysis is performed on the wafer failure map. The average density of each failure connected region is calculated using the failure density map. Pixels of failure connected regions with an average density higher than the average density of the entire image are set to zero to obtain the occlusion image. The occlusion image is then fed into a trained multi-label classifier for inference to obtain the occlusion probability vector. The difference between the occlusion probability vector and the initial probability vector is calculated to obtain the suppression recovery amount.

[0058] In this embodiment, based on the known density map and the average density of the whole map, it is necessary to identify which failed connected domains dominate the features of the whole map from the perspective of spatial structure, and remove their suppression effect on secondary failures by setting them to zero.

[0059] Perform 8-connected-part analysis on the binary failure pixel map of the wafer failure map to extract all failure connected parts, denoted as . , This represents the total number of failed connected components in this wafer failure diagram, and each failed connected component... From its pixel set Full description. 8-connectivity analysis groups adjacent failure pixels in the horizontal, vertical, and diagonal directions into the same connected region. It is suitable for capturing failure regions of various geometric shapes in wafer failure maps, including different shapes such as rings, clumps, and lines.

[0060] The average density of each failed connected component is calculated using the failure density map. The average density satisfies the following relationship:

[0061] In the formula, For failed connected components The average density is dimensionless. For the failed connected components A set of spatial cells with pixel overlap, i.e., satisfying the spatial cell condition. and All spatial cells where pixels intersect; This represents the number of elements in the set, expressed in units of individual elements. The first in the failure density diagram The cell density value of the space cell is derived from the calculation result of S102.

[0062] It should be noted that the average density integrates both the area information of the failed connected components and the local density: the more cells covered, the larger the failed area; the higher the average cell density value, the more continuous and compact the failure in that area. Compared to simply sorting by pixel area, the average density can accurately distinguish between large but sparse random scattered failures and medium-sized but highly dense central failures, the latter having a stronger dominant effect on the overall image features.

[0063] The occluded image is obtained by setting the pixels of the failed connected components whose average density is higher than the average density of the entire image to zero. Specifically, after determining the set of dominant failed connected components, the pixels corresponding to the dominant failed connected components are set to zero, and the occluded image satisfies the following relationship:

[0064] In the formula, To occlude coordinates in the image Pixel value at; It is the union of the pixel sets of all failed connected components in the dominant failed connected component set; The original wafer failure map in coordinates The pixel value at that location. Since the pixel value corresponding to the chip is 0 after the wafer failure map is normalized, setting the dominant failure connected region to zero is equivalent to setting it to a through state, which is consistent with the zeroing operation in the occlusion enhancement during the training phase.

[0065] When the average density of all failed connected components is not higher than the average density of the entire image, the pixels of the failed connected component with the highest average density are set to zero to obtain the occluded image.

[0066] It should be noted that the above fallback logic applies to scenarios with extremely low overall wafer yield, such as when the yield is below 50%. Large-area, uniformly distributed failures on the wafer cause the average density of all failed connected components to be lower than the average density of the entire graph. In this case, the solution degenerates into selecting the single failed connected component with the highest average density and setting it to zero. This ensures that the set of dominant failed connected components is not empty under any yield condition, guaranteeing the feasibility of the solution under boundary conditions.

[0067] For example, a wafer failure map has three failure connected regions: Edge-Ring type large area connected regions with an average density of 0.85, Center type medium area connected regions with an average density of 0.72, and Scratch type elongated connected regions with an average density of 0.18. The overall image density is 0.38. The Edge-Ring type connected regions and Center type connected regions have an average density higher than the overall image density. The pixels of both of them are set to zero, while the pixels of the Scratch type elongated connected regions are preserved in the occluded image.

[0068] After scaling the occluded image to 224×224 pixels, it is fed into a trained multi-label classifier for a second forward inference, yielding an occlusion probability vector. This occlusion probability vector is denoted as... ,in The first image obtained by inference from a trained multi-label classifier for the occluded image The predicted probability after occlusion for the failure mode is dimensionless. Both inferences use the exact same trained multi-label classifier parameters, requiring no additional training.

[0069] The difference between the occlusion probability vector and the initial probability vector is calculated to obtain the suppression recovery amount. The suppression recovery of the failure mode satisfies the following relationship:

[0070] In the formula, For the first The suppression recovery of the failure mode is dimensionless and its value range is [value range missing]. ; The occlusion probability vector is the first... Predicted probability of occlusion for failure modes; The first probability vector in the initial probability vector The probability of the existence of a failure mode is derived from the calculation results of S102.

[0071] Understandably, suppressing recovery volume The numerical value corresponds to three typical physical meanings: when When the value is significantly less than 0, it indicates that this type of failure is the dominant failure mode itself, which has been set to zero, and its predicted probability decreases significantly due to the occlusion of the dominant region; when When the value is greater than 0, it indicates that this type of failure is suppressed by the dominant failure connectivity in the original wafer failure map, and is a real secondary failure mode. The suppression recovery directly quantifies the depth of suppression; when When the value is close to 0, it indicates that the predicted probability of this type of failure is not affected by the dominant failure shading operation and does not exist in this wafer failure map.

[0072] By constructing occlusion images and performing a second inference, the secondary failure signals that were obscured in the original wafer failure image due to the large difference in area ratio are explicitly presented in the form of quantifiable suppression recovery amount, providing personalized recovery basis for each wafer failure image and each type of failure mode for subsequent adaptive multi-label decision.

[0073] S104: Build a matrix to determine the threshold and output the classification results under two conditions.

[0074] A suppression relation matrix is ​​constructed based on historical inference data. The suppression relation matrix is ​​multiplied by a preset safety coefficient to obtain an adaptive threshold. When the probability of a certain class in the initial probability vector is not lower than the preset probability threshold, or the corresponding suppression recovery exceeds the adaptive threshold, the final label of that class is set to 1, and the final multi-label classification result is output.

[0075] In this embodiment, after obtaining the initial probability vector and suppression recovery amount, a final decision needs to be made for each type of failure mode. Directly applying a fixed threshold to treat all category pairs equally fails to reflect the differences in suppression depth between different dominant failure categories and different secondary failure categories: the suppression depth of large-area dense failures on slender and sparse failures differs significantly physically from the suppression depth of central failures on random scattered failures due to varying degrees of area difference. Therefore, it is necessary to statistically analyze the typical suppression depth between each category pair from historical inference data accumulated in the production line, construct a suppression relationship matrix, and generate an adaptive threshold based on this matrix.

[0076] Constructing a suppression relationship matrix based on historical inference data includes: extracting the dominant failure category as the first from the historical inference data. And the first The set of historical samples of actual failure types is used, and the dominant failure type is the type with the highest probability in the initial probability vector of the corresponding historical sample. The mean value of the suppression recovery corresponding to each historical sample in the set of historical samples is calculated to obtain the suppression relationship matrix. Line number Column elements; suppression relation matrix Line number The column elements are positively correlated with the corresponding adaptive threshold, i.e., the [number]th [element]. Line number The larger the column element, the larger the corresponding adaptive threshold.

[0077] Specifically, construct Suppression Relationship Matrix , its first The element satisfies the following relation:

[0078] In the formula, For the suppression relation matrix, the first Line number Column elements, dimensionless; For historical inference data, the dominant failure category is the first. And the first The set of historical samples of actual failures, with the dominant failure category being the category with the highest probability in the initial probability vector of the corresponding historical samples; The number of samples in the historical sample set, expressed in units of individuals; For the first The first historical sample calculated in S103 Suppression recovery amount of failure modes. The larger the value, the higher the value. When class failure is the dominant failure, the first The deeper the average suppression of secondary failures, the more likely the triggering of the first... The threshold for the amount of suppression and restoration required for supplementary judgments should also be correspondingly higher.

[0079] Understandably, the suppression relationship matrix is ​​obtained entirely from the suppression recovery amount and historical label statistics produced in the historical inference data, without the need to introduce any new data or new algorithms, and it continues to converge and refine as production line data accumulates.

[0080] Constructing a suppression relation matrix based on historical inference data also includes: counting the number of historical samples in the historical sample set; when the number is less than the preset minimum number of samples, the first... Class-dominant failure on the first The adaptive threshold for class failure is set to the preset cold start threshold; when the number is not less than the preset minimum number of samples, the preset cold start threshold is switched to the threshold for suppressing the relation matrix. Line number The corresponding adaptive threshold is obtained by multiplying the column element by the preset safety factor.

[0081] It should be noted that the preferred reference value for the minimum preset sample size is 30, and the preferred reference value for the preset cold start threshold is 0.10, which is lower to ensure priority for recall during the early stages of production line data scarcity. When the number of samples in the historical sample set is less than 30, the suppression relation matrix is... The statistical results for the element are not yet reliable, and the corresponding adaptive threshold uses a preset cold start threshold of 0.10 as a conservative initial value; when the number of samples is not less than 30, it automatically switches to the suppression relation matrix. An adaptive threshold is calculated by multiplying the element with a preset safety factor, enabling a smooth transition from conservative fixed values ​​to data-driven refined values.

[0082] Multiplying the suppression relationship matrix by the preset safety factor yields the adaptive threshold, which satisfies the following relationship:

[0083] In the formula, For the first When class failure is the dominant failure, the first The adaptive threshold used for class failure determination is dimensionless. The preset safety factor is dimensionless, and the preferred reference value is 0.6; For the suppression relation matrix, the first Line number Column elements.

[0084] The preset safety factor is determined as follows: the recall rate of each failure mode in the validation set is not lower than the recall rate corresponding to the fixed threshold scheme. Within the range of values ​​that satisfy the constraint, the preset safety factor is obtained by searching with the goal of maximizing the average F1 score of the validation set.

[0085] Understandably, setting the preset safety factor to 0.6, that is, using 60% of the historical average suppression depth as the decision threshold, means that in a certain inference, the first... If the suppression recovery amount of a type of failure reaches 60% of the historical typical suppression depth, a supplementary decision is triggered to ensure that the suppression recovery amount of the vast majority of real minor failures can exceed the adaptive threshold. This improves the recall rate while preventing the false positive rate from being too high through constraints.

[0086] The dominant failure category index of the current wafer failure map is determined to be the category with the highest probability in the initial probability vector, denoted as . Combining the initial probability vector and the suppression recovery amount, a final multi-label decision is performed for each failure mode. The final multi-label classification result satisfies the following relationship:

[0087] In the formula, For the first The final determination label for the class of failure modes is dimensionless. ; The first probability vector in the initial probability vector The probability of the existence of a failure mode is derived from S102; The preset probability threshold is dimensionless, and the preferred reference value is 0.5. For the first The amount of suppression recovery for the failure mode comes from S103; dominant failure category For the first The adaptive threshold used for class failure determination; For logical OR operation, if any condition is met, then let .

[0088] It should be noted that this decision rule includes two complementary decision paths. The first path sets the final label of the class to 1 when the initial probability is not lower than a preset probability threshold, preserving the original decision-making ability of the multi-label classifier for dominant failures and independently occurring single failure modes, ensuring that the classification accuracy of the scheme is not affected in non-mixed scenarios. The second path sets the final label of the class to 1 when the suppression recovery exceeds an adaptive threshold, specifically for mixed failure scenarios: although the existence probability of secondary failures is lower than the preset probability threshold due to the significant difference in area proportion, their suppression recovery exceeds the adaptive threshold because the signal is desuppressed after the dominant failure is removed, thereby triggering supplementary judgment and correcting the missed judgment in mixed failure scenarios.

[0089] For example, continuing the scenario in S103: Edge-Ring and Center are the primary failures, Scratch is the secondary failure, and the primary failure category index is... This corresponds to the Edge-Ring class. Historical inference data shows that the element value of Edge-Ring to Scratch in the suppression relationship matrix is ​​0.25. Therefore, the adaptive threshold is... In the original inference, the probability of the Scratch class is 0.31, which is lower than the preset probability threshold of 0.5, so the first path is not triggered. After occluding the Edge-Ring and Center, the prediction probability of the Scratch class after occlusion increases to 0.53, and the suppression recovery amount is 0.22, which exceeds the adaptive threshold of 0.15, so the second path is triggered, the final label of the Scratch class is set to 1, and the missed detection is corrected.

[0090] After outputting the final multi-label classification result, the suppression recovery amount corresponding to the current wafer failure map and the final multi-label classification result are recorded to the historical inference data. According to the preset update cycle, the suppression relationship matrix is ​​re-statistically calculated based on the accumulated historical inference data, and the adaptive threshold is updated. The preferred reference value for the preset update cycle is once a quarter.

[0091] Understandably, the suppression relationship matrix is ​​recalculated quarterly based on accumulated historical inference data, so that the adaptive threshold continues to converge and refine with the accumulation of production line data, forming a data flywheel effect. The judgment accuracy automatically improves with the production line running time, without the need for manual intervention and recalibration.

[0092] By using a suppression relationship matrix to structurally precipitate the suppression patterns between different categories in the production line's historical experience, and combining this with a preset safety factor to generate an adaptive threshold for each type of failure mode, the final multi-label classification result is output through two complementary decision paths. This fundamentally eliminates the suppression effect of significant area ratio differences on the recall rate of minor failures, and enables accurate supplementary identification of minor failure modes in mixed failure scenarios, effectively reducing the directional bias of root cause analysis in mixed modes.

[0093] like Figure 2 As shown, the classification accuracy of the method in this embodiment is significantly higher than that of the traditional fixed threshold method across all eight failure modes. The most significant improvements are seen in the accuracy of the Scratch and Random classes, increasing from 0.68 and 0.71 to 0.80 and 0.84, respectively, representing relative improvements of 12% and 13%. This is because the present invention, through an occlusion-based re-inference mechanism, can effectively identify secondary failure modes suppressed by the dominant failure mode, while the traditional fixed threshold method easily misclassifies secondary failures as noise in mixed failure scenarios.

[0094] This figure directly verifies the technical effect of multi-label failure mode classification by occlusion re-inference and suppression signal recovery calculation. The ResNet-50 multi-label classifier in this embodiment, combined with adaptive decision threshold, improves the recognition accuracy of various failure modes by an average of 15.3%, meeting the industrial requirement of ≥85% automatic failure mode classification accuracy for semiconductor production lines.

[0095] like Figure 3 As shown, all diagonal elements are negative, indicating that the prediction probability of the dominant failure drops significantly after being occluded, verifying the effectiveness of the occlusion operation. All off-diagonal elements are positive, indicating that the prediction probability of secondary failures increases after the dominant failure is occluded, quantifying the suppression depth. Notably, the suppression amount of Edge-Ring on Scratch is significantly higher than that of Center on Random, because large-area dense failures occupy a stronger feature space for slender and sparse failures.

[0096] The figure directly illustrates the physical meaning of the constructed suppression relationship matrix, which reflects the category-specific suppression relationship between different failure mode pairs. This allows the decision threshold to be automatically adjusted according to the physical characteristics of the failure mode pairs, avoiding misjudgments caused by treating all category pairs equally with a fixed threshold.

Claims

1. A method for analyzing and classifying chip test failure modes based on artificial intelligence, characterized in that, Includes the following steps: Based on historical failure maps and multi-label vectors, a multi-label classifier is trained using occlusion enhancement to obtain a trained multi-label classifier; the wafer failure map to be classified is obtained and fed into the trained multi-label classifier for inference to obtain an initial probability vector; The wafer failure map is divided into uniform grids to construct a failure density map, and the average density of the entire map is calculated. Connectivity analysis is performed on the wafer failure map. The average density of each failure connected region is calculated using the failure density map. Pixels in failure connected regions with an average density higher than the average density of the entire image are set to zero to obtain the occlusion image. The occlusion image is fed into a trained multi-label classifier for inference to obtain the occlusion probability vector. The difference between the occlusion probability vector and the initial probability vector is calculated to obtain the suppression recovery amount. A suppression relationship matrix is ​​constructed based on historical inference data. The suppression relationship matrix is ​​multiplied by a preset safety coefficient to obtain an adaptive threshold. When the probability of a certain class in the initial probability vector is not lower than the preset probability threshold, or the corresponding suppression recovery amount exceeds the adaptive threshold, the final label of that class is set to 1, and the final multi-label classification result is output. Training a multi-label classifier using occlusion enhancement includes: in each training batch, randomly selecting a portion of training samples with a preset enhancement probability; performing connected component analysis on the selected training samples, randomly selecting one of the failed connected components and setting its pixels to zero to obtain enhanced samples, the labels of the enhanced samples being consistent with the original training samples; and inputting the enhanced samples together with the unselected training samples in the current batch to update the parameters of the multi-label classifier. Constructing a suppression relationship matrix based on historical inference data includes: extracting a set of historical samples from the historical inference data where the dominant failure category is type j and type k failures actually exist, with the dominant failure category being the category with the highest probability in the initial probability vector of the corresponding historical sample; calculating the mean of the suppression recovery amount corresponding to each historical sample in the historical sample set to obtain the element in the j-th row and k-th column of the suppression relationship matrix; the element in the j-th row and k-th column of the suppression relationship matrix is ​​positively correlated with the corresponding adaptive threshold, i.e., the larger the element in the j-th row and k-th column, the larger the corresponding adaptive threshold; and also includes: counting the number of historical samples in the historical sample set; when the number is less than the preset minimum number of samples, setting the adaptive threshold of type j dominant failure to type k failure as the preset cold start threshold; when the number is not less than the preset minimum number of samples, switching the preset cold start threshold to the product of the element in the j-th row and k-th column of the suppression relationship matrix and the preset safety coefficient to obtain the corresponding adaptive threshold.

2. The method for analyzing and classifying chip test failure modes based on artificial intelligence according to claim 1, characterized in that, Constructing the failure density map includes: dividing the wafer failure map into M×M spatial cells, where the value of M is chosen such that the average number of chips in each spatial cell is not less than a preset minimum number of chips; calculating the ratio of the number of failed pixels in each spatial cell to the total number of pixels in that spatial cell to obtain the cell density value of each spatial cell; the cell density values ​​of all spatial cells constitute the failure density map; and averaging the cell density values ​​of all spatial cells to obtain the average density value of the entire map.

3. The method for analyzing and classifying chip test failure modes based on artificial intelligence according to claim 1, characterized in that, When the average density of all failed connected components is not higher than the average density of the entire image, the pixels of the failed connected component with the highest average density are set to zero to obtain the occluded image.

4. The method for analyzing and classifying chip test failure modes based on artificial intelligence according to claim 1, characterized in that, The preset safety factor is determined as follows: the recall rate of each failure mode in the validation set is not lower than the recall rate corresponding to the fixed threshold scheme. Within the range of values ​​that satisfy the constraint, the preset safety factor is obtained by searching with the goal of maximizing the average F1 score of the validation set.

5. The method for analyzing and classifying chip test failure modes based on artificial intelligence according to claim 1, characterized in that, After outputting the final multi-label classification result, the following steps are also included: recording the suppression recovery amount corresponding to the current wafer failure map and the final multi-label classification result into the historical inference data; and re-statistically calculating the suppression relationship matrix based on the accumulated historical inference data according to the preset update cycle, and updating the adaptive threshold.

6. The method for analyzing and classifying chip test failure modes based on artificial intelligence according to claim 1, characterized in that, The multi-label classifier includes: a residual convolutional network, which extracts residual features layer by layer from the input wafer failure map and outputs a feature vector after global average pooling; a fully connected layer, which maps the feature vector to a K-dimensional output vector, where K is the total number of predefined failure mode categories; and K independent activation functions, which independently calculate each dimension of the K-dimensional output vector and output the existence probability of each failure mode. The existence probabilities of each failure mode are independent of each other and are not subject to the normalization constraints of other category probabilities.

7. The method for analyzing and classifying chip test failure modes based on artificial intelligence according to claim 6, characterized in that, The training loss used to train the multi-label classifier using occlusion enhancement is determined as follows: calculate the binary cross-entropy between the existence probability of each of the K failure mode categories and the corresponding multi-label vector components, and calculate the average of the binary cross-entropy of all categories to obtain the training loss. The parameters of the multi-label classifier are updated using the training loss as a supervision signal until the average F1 score on the validation set no longer improves.

Images