Background Image-based overlay (IBO) target is a type of overlay metrology method in semiconductor manufacturing. Focus optimization is a critical factor that determines the quality of the IBO target signals in IBO target measurement. To calculate the overlay, two target images are required. As the semiconductor process becomes complex, optimizing the IBO target focus for both targets becomes increasingly difficult. Aim We target to develop a deep-learning-based method that tackles two issues in IBO target focus optimization. Instead of choosing a compromised focus offset for both targets or measuring different targets sequentially, we train a generative model that outputs IBO target images with different focus offsets. We use the model for the IBO target focus optimization. Approach The proposed method consists of two steps. The first step is to train a model; conditional deblurring network (cDN) generates IBO target images with different camera focus offsets. Our model is based on the deep image deblurring methods using the generative models with auxiliary focus offsets. The second step is to optimize IBO target images with the trained model. We can generate an optimal target image that consists of IBO target images with the best focus for each layer. Results In experiments, we have compared different baseline models for cDN. The best model, which adopted DeblurGAN-v2, achieved a peak signal-to-noise-ratio of 49.04, a Fr'echet inception distance of 1.11, a mean squared error (MSE) of contrast precision of 0.0010, and an MSE of center-of-symmetry of 0.0672. The model can generate focus-adjusted IBO target images comparable to the ground truth. With the trained model, we optimized IBO target images. We achieved the improved target quality compared with the compromised best focus images. Conclusions We proposed a deep-learning-based methodology to solve the focus optimization problem for IBO target images. We demonstrated that our method can solve the IBO target optimization problem more efficiently than existing methods. In addition, we proved the applicability of deep learning in semiconductor manufacturing.
Hybrid bonding is a key enabling technology for advanced wafer-level packaging, providing high interconnect density and improved electrical performance. However, bonding interface quality remains a critical challenge, as void formation can severely impact yield and reliability. In this work, we investigate the influence of pre-bond wafer nanotopography (NT) and local warp on post-bond performance. Patterned Wafer Geometry (PWG) metrology is employed to characterize pre-bond NT and local warp, while acoustic scanning (CSAM) is used to evaluate post-bond void distribution. The results show that wafers with uniform NT profiles and minimal local warp achieve significantly improved bonding quality with reduced void formation. In contrast, wafers exhibiting large NT variation and asymmetric local warp are prone to severe void defects. Statistical analysis further indicates that within-site NT and local warp variation are strong indicators of bonding outcome. These findings demonstrate that effective control of pre-bond surface uniformity and local warp is essential for optimizing hybrid bonding processes and improving yield and reliability in wafer-level packaging.
In advanced 3D NAND flash memory fabrication, precise control of channel hole depth is critical for ensuring uniform electrical characteristics and reliable device performance. Recent studies have shown that increasing stack height amplifies etch non-uniformity and pattern-edge effects in channel holes. Variations in etch depth across the wafer can lead to "under-etch" or "over-etch," risking incomplete channel formation or punch-through failures, respectively. These challenges are exacerbated by the limited accessibility of deeply buried features, which restricts the capability of traditional cross-section SEM (X-SEM) and transmission electron microscopy (TEM) methods for high-volume metrology. Conventional techniques like X-SEM/TEM are destructive, time-consuming, and limited to localized sampling, making them unsuitable for large-scale process monitoring. To address these limitations, we present a non-destructive approach based on Critical Dimension Small-Angle X-ray Scattering (CD-SAXS) from KLA’s Axion® system. CD-SAXS has recently been identified as a promising next-generation inline metrology solution capable of characterizing 3D-NAND channel holes. This method leverages the sensitivity of scattering profiles to vertical structural dimensions, extracting depth information through an advanced modeling approach from KLA. This approach enables rapid process feedback, reduces metrology costs, and supports robust inline depth control in 3D NAND manufacturing.
This paper presents a comprehensive comparative evaluation of three advanced object detection configurations-YOLOv11, RT-DETR, and a hybrid ensemble-for automated weed detection in chilli (Capsicum annuum L.) fields under real-world agricultural conditions. A domain-specific dataset collected over an extended period was curated to capture diverse illumination, background complexity, and crop-weed growth stages. Each model was trained and validated to assess detection accuracy, precision, Inference Speed, bounding-box quality, and real-time performance. YOLOv11 delivered strong baseline performance with high accuracy and rapid inference, making it suitable for real-time field applications. RT-DETR demonstrated superior localization precision and robustness at higher IoU thresholds due to its transformer-based architecture, though with relatively lower inference speed. The proposed hybrid ensemble integrating YOLOv11 and RT-DETR achieved the most balanced and stable performance across all evaluation metrics, significantly reducing both false positives and false negatives compared to the individual models. These findings highlight the trade-offs between computational efficiency and detection robustness, illustrating that hybrid deep learning frameworks can effectively blend speed and accuracy for reliable, real-time weed detection. The study underscores the potential of ensemble-based architectures as optimal solutions for scalable precision agriculture applications.
A new design approach on 4H-SiC material is ongoing to improve the electrical performance of devices. As seen in silicon devices, multi-epitaxial growth enhances performance by reducing on-resistance (R on ). However, devices built on SiC face several challenges due to the very low dopant diffusion (e.g. phosphorus and aluminum) and defect evolution during the epitaxial growth. Monitoring defects like prismatic faults, stacking faults, partial dislocations, and micropipes, especially after regrowth, is essential to assess their impact on device performance. Defects with high killer ratio must be closely tracked to understand evolution thereof. In this work, we will show a method for early-stage process characterization and defect root-cause identification through sensitive inspections, effective reviews, and accurate defect classification to detect critical defects in 4H-SiC material when more than one epitaxial step is considered.
