Chinese Journal of Lasers, Volume. 51, Issue 21, 2109002(2024)
Wide Field‐of‐View Multiscale Noncontact Photoacoustic Intelligent Defect‐Detection Algorithm
During the fabrication of flip chips, challenges to production yield and longevity arise owing to preparation defects, including delamination, missing solder bumps, and cracks. These defects typically manifest at dimensions ranging from the submillimeter to micron scale and are characterized by pronounced randomness and broad distribution. Consequently, comprehensive defect detection across these multiscale dimensions facilitates the early identification and removal of flawed chips, thereby enhancing both the production yield and long-term operational stability. Whereas existing nondestructive testing methodologies such as ultrasonics, laser ultrasonics, X-ray computed tomography, pulsed phase thermography, and conventional photoacoustics partially satisfy the requirements of flip-chip detection, they exhibit certain challenges. These challenges include sample contamination, detection-speed limitations imposed by the average sampling procedures, potential risks associated with ionizing radiation, and susceptibility to environmental effects. Hence, this study introduces an intelligent defect-detection methodology based on non-interferometric noncontact photoacoustic microscopy (NINC-PAM). This approach is designed to achieve accurate and extensive detections of preparation defects across varying dimensions within flip chips. By offering a feasible technical solution for inline nondestructive defect detection during the fabrication process of flip chips, this methodology is promising for substantially improving both the production yield and operational lifespan of flip chips.
First, we established an NINC-PAM system based on elasto-optical theory and autonomously developed an optical?mechanical joint scanning imaging mode for the wide-field-of-view (FOV) imaging of flip-chip samples. Second, by leveraging the NINC-PAM system, we introduced a multiscale defect-detection algorithm named Chip-YOLO to identify preparation defects of varying sizes within flip chips. This algorithm enhances the original YOLOv8 architecture by sequentially incorporating a small-object detection (SOD) layer, large separable kernel attention (LSKA) module, and reparameterized generalized feature pyramid network (RepGFPN) to optimize the detection accuracy for defects of different sizes in flip chips. Third, the flip-chip samples were prepared in the following sequence: spin coating, dehydration, photolithography, development, evaporation, and lift-off. Delamination defects of various sizes were introduced into the samples via ultrasonic cleaning. Fourth, by employing the optical?mechanical joint scanning mode of the NINC-PAM system, the samples were subjected to multiple wide-FOV scans, thus resulting in a dataset containing 29706 defects of different sizes. Fifth, by utilizing a method based on absolute-scale target definition, the established dataset was statistically analyzed and quantified to classify defects of different sizes, thereby validating the rationality of the dataset.
The proposed wide-FOV multiscale intelligent defect-detection method, which is based on NINC-PAM, achieves the precise identification and localization of delamination defects across varying sizes within flip-chip samples on an extensive scale. Initially, by leveraging the dataset formulated from the NINC-PAM imaging results, the Chip-YOLO algorithm demonstrates a multiscale average accuracy (AP) of 60.1% under 12.4 MB of parameters and 39.8 GFLOPs of computation amount. This performance surpasses those of other classical detection algorithms, including YOLOv3, YOLOX, YOLOv7, and Faster R-CNN, in terms of both accuracy metrics and computational efficiency (Table 3). Subsequent ablation experimental findings reveal that the incorporation of the SOD layer, LSKA module, and RepGFPN into the foundational YOLOv8 architecture increases the multiscale AP of Chip-YOLO for delamination defects. Remarkably, without significantly increasing the model parameters or computational demand, Chip-YOLO achieves 3.3 percentage points enhancement in the multiscale AP relative to YOLOv8 (Table 4). More importantly, the experiment of wide-FOV multiscale intelligent detection based on the NINC-PAM system accomplished the intelligent detection of defects of various sizes in flip-chip samples exceeding 1 mm×1 mm in just 23 s. The detection accuracy of the proposed method demonstrates its more accurate defect-detection ability compared with other classical intelligent detection methods, thereby substantiating its superior performance in wide-FOV multiscale defect detection (Fig. 7). Moreover, the performance statistics show that Chip-YOLO offers more accurate and faster defect detections compared with other one- and two-stage algorithms (Table 5). In the current wide-FOV multiscale intelligent detection experiments, the time consumption of a single mechanical scan is on the order of hundreds of milliseconds, whereas that of a single optical scan is approximately 2 s, thus indicating that the experimental time is primarily governed by the galvanometer scanning process. Using a laser with megahertz repetition rates can increase the scanning frequency of the galvanometer for real-time imaging. Meanwhile, although Chip-YOLO offers better multiscale defect detection, it introduces additional parameters and computational overheads, thus prolonging the detection time. In the future, efficient CNN (convolutional neural networks) building blocks can be introduced to reduce the network’s parameter count and computational load for improving the detection speed. Furthermore, wide-FOV intelligent detection images are currently obtained by manually stitching multiple optical scan images after performing an experiment. Developing the appropriate intelligent stitching methods will allow real-time image stitching to be achieved during scanning and detection.
This study proposes a wide-FOV, multiscale intelligent defect-detection methodology based on the NINC-PAM system that is customized for the precise identification and intelligent localization of preparation defects of varying sizes in flip chips. This approach integrates the NINC-PAM system and operates in an optical–mechanical joint scanning mode with the proposed Chip-YOLO multiscale defect-detection algorithm to achieve accurate detections of delamination defects of different sizes within a wide FOV. On the established dataset of preparation defects in flip chips, Chip-YOLO achieves a multi-scale AP of 60.1%, surpassing the multiscale AP of other classical algorithms. Further ablation experimental results confirm that Chip-YOLO exhibits a higher multiscale AP compared with the original YOLOv8. More importantly, intelligent defect-detection experimental results prove that the proposed method can achieve accurate wide-FOV multiscale detection without increasing system latency. Similarly, performance statistics confirm that Chip-YOLO achieves more accurate and faster detections compared with other one- and two-stage algorithms. The proposed wide-FOV multiscale intelligent defect-detection method based on the NINC-PAM system demonstrates significant potential for online defect detection in flip chips.
Get Citation
Copy Citation Text
Jijing Chen, Yihan Pi, Yixuan Pang, Hao Zhang, Kaixuan Ding, Ying Long, Jiao Li, Zhen Tian. Wide Field‐of‐View Multiscale Noncontact Photoacoustic Intelligent Defect‐Detection Algorithm[J]. Chinese Journal of Lasers, 2024, 51(21): 2109002
Category: holography and information processing
Received: May. 15, 2024
Accepted: Jun. 17, 2024
Published Online: Oct. 31, 2024
The Author Email: Li Jiao (jiaoli@tju.edu.cn), Tian Zhen (tianzhen@tju.edu.cn)
CSTR:32183.14.CJL240877