A multi-agent bidirectional closed-loop collaborative wafer defect detection system and method

Abstract

The application discloses a kind of multi-agent bidirectional closed-loop collaborative wafer defect detection system and method, global screening agent is used to carry out global scanning to wafer defect image, and generates defect candidate area and defect probability heat map;Local precision inspection agent is extracted to subpixel level defect geometric parameter and fine-grained semantic feature through high-resolution imaging and fine detection to all defect candidate areas;Defect classification agent parallelly deploys multiple differentiated backbone networks as base classifier, respectively classifies the output results of local precision inspection agent, and obtains defect category by weighted average of the classification results of multiple differentiated backbone networks;Closed-loop error correction agent carries out confidence, size and feature consistency evaluation to the output results of defect classification agent, and outputs final result if evaluation is qualified, and re-inspects the defect candidate area if evaluation is unqualified.The application fundamentally suppresses multi-stage detection error accumulation, and realizes mutual compatibility of ultra-high detection precision and production line beat.

Description

Technical Field

[0001] This invention relates to the field of wafer defect detection technology, and in particular to a multi-agent bidirectional closed-loop collaborative wafer defect detection system and method. Background Technology

[0002] Wafer defect detection covers the entire process flow, including photolithography, etching, and deposition. Mainstream technologies include automated optical inspection (AOI), electron beam inspection (EBI), and scattering inspection. To balance inspection speed and sensitivity, the industry commonly employs a unidirectional cascaded inspection architecture: the front-end unit uses lightweight algorithms to perform rapid initial screening of the entire wafer, filtering out the vast majority of defect-free areas, and only sending suspected defective areas to the subsequent high-precision unit for verification, ultimately outputting the defect classification results. This architecture can improve inspection sensitivity while ensuring production line cycle time and is the standard link topology widely used in mainstream wafer inspection equipment worldwide.

[0003] Mainstream unidirectional cascaded inspection architectures have inherent error lower bounds, failing to eliminate error accumulation issues. Current AOI inspection employs a unidirectional serial link of "initial screening-precision inspection-classification," where data is only transmitted unidirectionally from the front stage to the back stage, lacking a reverse feedback channel. Error samples from the front stage continuously flow downstream, and the subsequent error correction has a performance limit, amplifying false detection errors at each stage. Secondly, existing inspection systems lack online real-time quality verification and targeted re-inspection mechanisms, making it difficult to balance accuracy and production line cycle time. Industry multi-agent systems are only used for back-end processes such as yield analysis and equipment scheduling, without being embedded in the main link for real-time error correction. Summary of the Invention

[0004] To overcome the aforementioned problems in the prior art, this invention proposes a multi-agent bidirectional closed-loop collaborative wafer defect detection system and method.

[0005] The technical solution adopted by this invention to solve its technical problem is: a multi-agent bidirectional closed-loop collaborative wafer defect detection system, including a global screening agent, a local fine inspection agent, a defect classification agent, and a closed-loop error correction agent. The global screening agent is used to perform a full-domain scan of the wafer defect image to generate defect candidate regions and defect probability heatmaps. The local fine inspection agent extracts sub-pixel-level defect geometric parameters and fine-grained semantic features from all defect candidate regions generated by the global screening agent through high-resolution imaging and fine detection. The defect classification agent deploys multiple differentiated backbone networks in parallel as base classifiers to classify the results output by the local fine inspection agent, and performs a weighted average of the classification results of the multiple differentiated backbone networks to obtain the defect category. The closed-loop error correction agent evaluates the confidence, size, and feature consistency of the output results of the defect classification agent. If the evaluation is qualified, the final result is output; if the evaluation is unqualified, the defect candidate region is re-inspected.

[0006] The aforementioned multi-agent bidirectional closed-loop collaborative wafer defect detection system uses a MobileNetV3-Large backbone network and a P3~P7 layer FPN multi-scale structure to scan candidate regions.

[0007] The aforementioned multi-agent bidirectional closed-loop collaborative wafer defect detection system integrates super-resolution reconstruction and cascaded mask detection networks to simultaneously acquire sub-pixel dimensions and high-precision fine-grained features of defects.

[0008] The aforementioned multi-agent bidirectional closed-loop collaborative wafer defect detection system uses a local precision detection agent for super-resolution reconstruction via Real-ESRGAN, feature extraction via ResNet-101 and FPN, cascaded detection via Cascade R-CNN, and pixel mask output via a Mask branch.

[0009] The aforementioned multi-agent bidirectional closed-loop collaborative wafer defect detection system has four types of differentiated backbone networks deployed in parallel by the defect classification agents: ResNet-152, ResNeSt-101, ResNeSt-101 and CBAM attention, and ResNeXt-101.

[0010] The aforementioned multi-agent bidirectional closed-loop collaborative wafer defect detection system uses high-resolution images and fine-grained feature vectors output by the local fine-detection agent as input for each differentiated backbone network of the defect classification agent. Weights are assigned to each differentiated backbone network, and the prediction results are fused by weighted averaging and voting. The fusion probability is obtained after weighted averaging. If the category with the highest probability receives votes from at least 3 differentiated backbone networks, the category is adopted; otherwise, a re-detection is triggered. The confidence vector and the highest confidence are output by temperature scaling calibration.

[0011] The aforementioned multi-agent bidirectional closed-loop collaborative wafer defect detection system includes a closed-loop error correction agent that sets a threshold vector. As a criterion for judgment, among them The confidence threshold is dynamically set based on the defect category. This is a size threshold, set according to the process node; This is the feature consistency threshold, which is a fixed value. Similarity is calculated by aligning the coarse and fine feature dimensions using the projection matrix. ,in, For the projection matrix, It is characterized by coarse-grained properties. It is a fine-grained feature; Relying on trigger function Determine if a re-detection is needed. If any condition is met, trigger a re-detection of that area, and simultaneously output a re-detection command and parameter adjustment plan. For defect detection confidence level. For the defect geometry, For indicator functions, For similarity.

[0012] A multi-agent bidirectional closed-loop collaborative wafer defect detection method, based on the aforementioned multi-agent bidirectional closed-loop collaborative wafer defect detection system, includes the following steps: Step 1: Input the wafer defect image into the global screening agent. The global screening agent performs a full-domain scan of the wafer defect image and outputs a set of candidate boxes with coordinates, defect probability and coarse-grained features. Step 2: The output results of Step 1 are processed by a local fine inspection agent for high-resolution imaging and fine detection to calculate the defect area and the equivalent diameter of the defect, thereby obtaining fine-grained features. Step 3: The defect classification agent takes the high-resolution image and fine-grained feature vector output from Step 2 as input, assigns different weights to different base classifiers, and uses weighted averaging plus voting to fuse the prediction results, outputting the defect category and confidence level. Step 4: Set the threshold vector for the closed-loop error correction agent. As a criterion for judgment, among them The confidence threshold is dynamically set based on the defect category. This is a size threshold, set according to the process node; The feature consistency threshold is a fixed value; a trigger function is calculated for each defect. When any condition is met, a re-examination of the candidate region is triggered, and a re-examination instruction and parameter adjustment scheme are output.

[0013] The beneficial effects of this invention are as follows: It establishes a two-way information flow link of four intelligent agents: global screening, local fine inspection, defect classification, and closed-loop error correction. This breaks the error lower bound of the unidirectional cascaded architecture at the topological level. Through reverse re-inspection, it achieves proactive recall of missed defects and correction of false positives, fundamentally suppressing the accumulation of multi-level detection errors. It constructs an online real-time quality self-assessment and triggered local re-inspection mechanism, establishing a triple threshold criterion based on the closed-loop error correction intelligent agent. This automatically identifies defect areas with insufficient reliability and performs targeted secondary inspection, while simultaneously and dynamically adjusting the detection parameters of each intelligent agent. Only a small number of non-conforming areas are re-inspected, and a maximum number of iterations is set to control the time consumption. This reduces missed detection errors with a slight increase in detection time, achieving compatibility between ultra-high detection accuracy and production line cycle time. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the system architecture of the present invention; Figure 2 This is a flowchart of the global screening intelligent agent workflow of the present invention; Figure 3 This is a flowchart of the local precision detection intelligent agent of the present invention; Figure 4 This is a flowchart of the defect classification intelligent agent of the present invention; Figure 5 This is a flowchart of the closed-loop error correction intelligent agent feedback process of the present invention. Detailed Implementation

[0015] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0016] like Figure 1 As shown, this embodiment discloses a multi-agent bidirectional closed-loop collaborative wafer defect detection system, including a global screening agent, a local fine inspection agent, a defect classification agent, and a closed-loop error correction agent. The four agents form a closed-loop collaborative detection through bidirectional information flow.

[0017] The working principle of the global screening intelligent agent is as follows: Figure 2 As shown, a MobileNetV3-Large backbone network is used, coupled with a multi-scale FPN structure of P3~P7 layers, corresponding to receptive fields ranging from 32×32 to 512×512 pixels. Candidate regions with anchor box sizes [16, 32, 64, 128, 256] and aspect ratios [0.5, 1, 2] are generated using the RPN. Pre-training is performed using the WM-811K dataset and proprietary production line data (totaling over 500,000 labeled wafer images), followed by fine-tuning with actual defect images from the production line. The loss function is a combination of Focal Loss and Smooth L1 Loss, outputting a set of candidate boxes. Each bounding box carries coordinates, defect probability, and 256-dimensional coarse-grained features. This module enables rapid, low-resolution scanning of the entire wafer defect image, generating candidate defect regions and defect probability heatmaps. It innovatively employs a lightweight, multi-scale detection scheme, balancing full-area scanning speed with coarse defect recall capability, efficiently pinpointing the locations of all potential defects.

[0018] The workflow of the local precision inspection intelligent agent is as follows: Figure 3As shown, its function is to extract sub-pixel-level defect geometric parameters and fine-grained semantic features from each candidate region output in the previous step through high-resolution imaging and fine detection. First, Real-ESRGAN is used to perform a 4x super-resolution on the candidate regions. Then, a Cascade R-CNN with three levels of IoU thresholds of 0.5 / 0.6 / 0.7 is built using ResNet-101+FPN as the backbone network. A mask is output using the Mask branch, and the defect area A (number of pixels × pixel physical size) is calculated. 2 ), through formula Calculate the equivalent diameter of the defect and extract 512-dimensional fine-grained features from the RoIAlign layer of Cascade R-CNN. This module innovatively integrates super-resolution reconstruction and cascaded mask detection networks to simultaneously acquire sub-pixel size of defects and high-precision fine-grained features, thus compensating for the insufficient accuracy of coarse screening images.

[0019] The workflow of the defect classification agent is as follows: Figure 4 As shown, four differentiated backbone networks are deployed in parallel as base classifiers: ResNet-152, ResNeSt-101, ResNeSt-101+CBAM attention, and ResNeXt-101 (32×8d). Each network takes the high-resolution image and fine-grained feature vector from the fine-detection output as input. Weights of 0.25, 0.28, 0.25, and 0.22 are assigned based on the F1 score of each model on the validation set. A weighted average plus voting fusion of prediction results is used, and the fusion probability is obtained after weighted averaging. If the highest probability category receives votes from at least three base classifiers, that category is ultimately adopted; otherwise, a re-detection is triggered. The probabilities are then calibrated by temperature scaling, and a confidence vector is output. With the highest confidence level This module innovatively adopts a multi-model integration and fusion configuration reliability calibration strategy, which effectively reduces model prediction bias and significantly improves the reliability of various wafer defect classifications.

[0020] Setting threshold vector for closed-loop error correction agent As a criterion for evaluation, among them: (Confidence threshold): Dynamically set according to the defect category, with a typical range of 0.7~0.95. A higher value is used for critical defects (such as bridging) and a lower value is used for non-critical defects.

[0021] (Size threshold): Set according to the process node, 0.5μm for 7nm process, 1.0μm for 14nm process, and 2.0μm for 28nm process.

[0022] (Feature consistency threshold): Fixed at 0.75, which has been experimentally verified as the optimal balance point.

[0023] via projection matrix Align the coarse and fine feature dimensions and calculate similarity. Relying on trigger functions Determine if a re-detection is needed. If any condition is met, trigger a re-detection of the area. Simultaneously output the re-detection command (including the re-detection area ID) and parameter adjustment scheme (resolution enhancement coefficient, sensitivity threshold increment, etc.).

[0024] For a two-way closed-loop feedback process for multi-agent collaboration, such as Figure 5 As shown. It includes the following steps: Step 1: Input the wafer defect image into the global screening agent. The global screening agent performs a full-domain scan of the wafer defect image and outputs a set of candidate boxes with coordinates, defect probability and coarse-grained features. Step 2: The output results of Step 1 are processed by a local fine inspection agent for high-resolution imaging and fine detection to calculate the defect area and the equivalent diameter of the defect, thereby obtaining fine-grained features. Step 3: The defect classification agent takes the high-resolution image and fine-grained feature vector output from Step 2 as input, assigns different weights to different base classifiers, and uses weighted averaging plus voting to fuse the prediction results, outputting the defect category and confidence level. Step 4: Set the threshold vector for the closed-loop error correction agent. As a criterion for judgment, among them The confidence threshold is dynamically set based on the defect category. This is a size threshold, set according to the process node; The feature consistency threshold is a fixed value; a trigger function is calculated for each defect. When any condition is met, a re-examination of the candidate region is triggered, and a re-examination instruction and parameter adjustment scheme are output.

[0025] During retesting, press Update detection parameters, among which This includes: RPN score threshold (initially 0.7), NMS threshold (initially 0.5), super-reconstruction factor (initially 4x), and classification confidence threshold (initially 0.5). Decreasing with the number of iterations ( To avoid over-adjustment. The attenuation coefficient is among them. This embodiment innovatively designs an adaptive closed-loop error correction logic based on a three-indicator joint judgment, enabling dynamic parameter optimization and local iterative re-examination, constructing a complete bidirectional feedback detection link, and continuously suppressing missed detections and false detections.

[0026] This invention employs a global screening agent (lightweight coarse recall) → a local fine inspection agent (super-resolution reconstruction + sub-pixel contour extraction) → a defect classification agent (ensemble learning + confidence calibration) → a closed-loop error correction agent (quality monitoring and re-inspection decision-making), with bidirectional information flow between the agents. A closed-loop error correction agent is added to evaluate detection quality in real time, triggering reverse re-inspection for regions where confidence, size, or feature consistency does not meet thresholds, and dynamically adjusting preceding parameters to gradually eliminate errors through multiple iterations. The shallow features from the coarse screening stage are compared with the deep semantic features from the fine inspection stage using cosine similarity, serving as the quantitative criterion for triggering reverse re-inspection. Targeted re-inspection is performed only on local regions that meet the triggering conditions, slightly increasing the overall detection time while balancing detection accuracy and efficiency.

[0027] The above embodiments are merely exemplary embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art can make various modifications or equivalent substitutions to the present invention within its scope and protection, and such modifications or equivalent substitutions should also be considered to fall within the protection scope of the present invention.

Claims

1. A multi-agent bidirectional closed-loop collaborative wafer defect detection system, characterized in that, It includes a global screening agent, a local fine inspection agent, a defect classification agent, and a closed-loop error correction agent. The global screening agent is used to perform a full-domain scan of the wafer defect image and generate defect candidate regions and defect probability heatmaps. The local fine inspection agent extracts sub-pixel level defect geometric parameters and fine-grained semantic features from all defect candidate regions generated by the global screening agent through high-resolution imaging and fine detection. The defect classification agent deploys multiple differentiated backbone networks in parallel as base classifiers to classify the results output by the local fine inspection agent. The classification results of the multiple differentiated backbone networks are weighted and averaged to obtain the defect category. The closed-loop error correction agent evaluates the confidence, size, and feature consistency of the output results of the defect classification agent. If the evaluation is qualified, the final result is output. If the evaluation is unqualified, the candidate defect region is re-inspected.

2. The multi-agent bidirectional closed-loop collaborative wafer defect detection system according to claim 1, characterized in that, The global screening agent uses the MobileNetV3-Large backbone network, coupled with a P3~P7 layer FPN multi-scale structure, to scan candidate regions.

3. The multi-agent bidirectional closed-loop collaborative wafer defect detection system according to claim 1, characterized in that, The local fine inspection agent integrates super-resolution reconstruction and cascaded mask detection networks to simultaneously acquire sub-pixel size of defects and high-precision fine-grained features.

4. The multi-agent bidirectional closed-loop collaborative wafer defect detection system according to claim 3, characterized in that, The local fine-detection agent performs super-resolution reconstruction using Real-ESRGAN, extracts features using ResNet-101 and FPN, performs cascaded detection using Cascade R-CNN, and outputs pixel masks through the Mask branch.

5. The multi-agent bidirectional closed-loop collaborative wafer defect detection system according to claim 1, characterized in that, There are four types of differentiated backbone networks deployed in parallel by the defect classification agent: ResNet-152, ResNeSt-101, ResNeSt-101 with CBAM attention, and ResNeXt-101.

6. The multi-agent bidirectional closed-loop collaborative wafer defect detection system according to claim 5, characterized in that, Each differentiated backbone network of the defect classification agent takes the high-resolution image and fine-grained feature vector output by the local fine-detection agent as input, assigns weights to each differentiated backbone network, and uses weighted averaging plus voting to fuse the prediction results. The fusion probability is obtained after weighted averaging. If the category with the highest probability receives at least 3 votes from the differentiated backbone networks, the category is adopted; otherwise, a re-detection is triggered. The confidence vector and the highest confidence are output by temperature scaling calibration.

7. The multi-agent bidirectional closed-loop collaborative wafer defect detection system according to claim 1, characterized in that, The closed-loop error correction agent sets a threshold vector. As a criterion for judgment, among them The confidence threshold is dynamically set based on the defect category. This is a size threshold, set according to the process node; This is the feature consistency threshold, which is a fixed value. Similarity is calculated by aligning the coarse and fine feature dimensions using the projection matrix. ,in, For the projection matrix, It is characterized by coarse-grained properties. It is a fine-grained feature; Relying on trigger function Determine if a re-detection is needed. If any condition is met, trigger a re-detection of that area, and simultaneously output a re-detection command and parameter adjustment plan. For defect detection confidence level. For the defect geometry, For indicator functions, For similarity.

8. A multi-agent bidirectional closed-loop collaborative wafer defect detection method, characterized in that, A multi-agent bidirectional closed-loop collaborative wafer defect detection system according to any one of claims 1-7 includes the following steps: Step 1: Input the wafer defect image into the global screening agent. The global screening agent performs a full-domain scan of the wafer defect image and outputs a set of candidate boxes with coordinates, defect probability and coarse-grained features. Step 2: The output results of Step 1 are processed by a local fine inspection agent for high-resolution imaging and fine detection to calculate the defect area and the equivalent diameter of the defect, thereby obtaining fine-grained features. Step 3: The defect classification agent takes the high-resolution image and fine-grained feature vector output from Step 2 as input, assigns different weights to different base classifiers, and uses weighted averaging plus voting to fuse the prediction results, outputting the defect category and confidence level. Step 4: Set the threshold vector for the closed-loop error correction agent. As a criterion for judgment, among them The confidence threshold is dynamically set based on the defect category. This is a size threshold, set according to the process node; The feature consistency threshold is a fixed value; a trigger function is calculated for each defect. When any condition is met, a re-examination of the candidate region is triggered, and a re-examination instruction and parameter adjustment scheme are output.

Images