SEMI-SuperYOLO-NAS：高NA EUVL半导体缺陷检测的增强方法

版权申诉

155 浏览量更新于2024-06-13 收藏 16.34MB PDF 举报

"基于SEMI-SuperYOLO-NAS的高NA EUVL半导体缺陷检测改进" 这篇研究论文探讨了在高数值孔径极紫外光刻（High-NA EUVL）技术中，如何改善半导体缺陷检测的问题。随着摩尔定律的持续发展，半导体行业正逐步采用高NA EUVL技术，以实现更小的间距。然而，这种技术的低焦深对大规模生产带来了挑战，如使用更薄的光刻胶和新型底层/硬掩模，可能导致信噪比（SNR）下降，从而使缺陷检测变得更加复杂。为解决这一问题，研究者提出了一种名为SEMI-SuperYOLO-NAS的半监督自适应缺陷分类检测（ADCD）框架，它是基于YOLO-NAS架构的扩展。SEMI-SuperYOLO-NAS引入了一个超分辨率（SR）辅助分支，该分支协助主干网络学习高分辨率特征，特别是在检测纳米级缺陷时。通过这个辅助分支，可以自动生成放大的图像，以适应不同图像分辨率的缺陷检测，而无需进行额外的显式训练。此外，研究中还研究了增强数据策略，以生成更多样化和真实的训练数据集，进一步提升模型的性能。这种方法在两个不同的工厂数据集上进行了测试，这些数据集源自两个不同的工艺，并由两种不同的成像工具捕获，显示了方法的通用性和有效性。 YOLO（You Only Look Once）是一种流行的实时目标检测算法，而YOLO-NAS是在YOLO基础上结合神经架构搜索（NAS）进行优化的版本。在SEMI-SuperYOLO-NAS中，NAS被用来自动寻找最优的网络结构，以提升缺陷检测的准确性和效率。总结来说，这项研究展示了如何利用机器学习，特别是基于视觉的深度学习算法，如YOLO和NAS，来改进高NA EUVL技术下的半导体缺陷检测。通过集成超分辨率技术并优化数据增强策略，SEMI-SuperYOLO-NAS为在各种图像分辨率下检测纳米级缺陷提供了有效解决方案，有助于半导体行业的质量控制和工艺优化。

model for nano-scale defect inspection and classiﬁcation. An unsupervised machine learning model was applied

to reduce false-positive defects arising from SEM noise. Dey et al. also benchmarked another ensemble defect de-

tection framework based on the state-of-the-art one-stage object detection model YOLOv5,

which outperformed

the previous ResNet-based ensemble framework.

A data-centric approach was also demonstrated,

showing a way to consensually ensemble labels from dif-

ferent anonymous labelers for the same dataset, requiring minimal intervention from experts for defect labeling

expectations. The ﬁnal combined label, along with expert-aware post-processing data, yielded the highest test

mAP of 0.919 with the state-of-the-art YOLOv8 model. The concept of pseudo-label prediction and model en-

sembles from diﬀerent labelling partitions to avoid duplicate labeling work was proposed. Cheon et al.

utilized

a Convolutional Neural Network (CNN) along with a k-NN method to classify wafer surface defects from SEM

images. The ﬁnal activation layer outputs were used as features for k-Nearest Neighbors (k-NN), but with a

threshold for clustering, allowing the algorithm to identify unknown defect classes. However, when a new class

is discovered, the CNN needs retraining to ensure consistent high performance over time. To address this issue,

the authors proposed the idea of incremental training. In previous study,

a noise characterization for future

denoising algorithms demonstrated that as scanning speed increased, a shift in noise property from a Gaussian

to a Gamma distribution was observed. In recent years, CD-SEM tool vendors have also incorporated machine

learning to enhance tool eﬃciency.

2, 3

The idea of ML-based super-resolution defect inspection was proposed,

where a CNN model takes the database of designed features and CD-SEM images as input and generates the

pixel distribution for defect prediction. Their model utilized non-defect regions to train the algorithm, generat-

ing the pixel luminosity distribution for the given database. If the pixel brightness values of the original image

signiﬁcantly deviated from the predicted pixel distribution center, the pixel was classiﬁed as defective. This ap-

proach demonstrated the potential for identifying bridge defects and necking defects even in lower pixel resolution

images of 4nm/pixel. However, in this method, knowledge of the design database (mask layout) is necessary,

making automated defect type classiﬁcation challenging. Nonetheless, their research highlights the advantages

of massive metrology, showcasing how a FOV of 80×80 microns can facilitate automatic defect identiﬁcation

through machine learning. Thus, the development of a super-resolution ML method capable of both identifying

and classifying defects without requiring knowledge of the mask design remains a crucial missing component.

3. METHODOLOGY

In this section, we provide a brief introduction to our training datasets, the proposed data augmentation strategy,

an overview of the architecture of our proposed SEMI-SuperYOLO-NAS, and the strategy for preparing the

dataset of High-resolution/Low-resolution (HR/LR) image pairs to train our SR-branch assisted defect detector

model.

3.1 Dataset

Two Line-Space (LS) pattern datasets were utilized. The ﬁrst dataset, SEM-ADI, comprises 1324 images captured

with the CD-SEM tool during After-Development-Inspection (ADI). The second dataset, EDR-AEI, comprises

527 images captured with the EDR tool during After-Etching-Inspection (AEI). Each image in these datasets

contains at least one defect. SEM-ADI and EDR-AEI have original pixel resolutions of 1024×1024 and 480×480,

respectively. SEM-ADI encompasses 5 diﬀerent defect types, while EDR-AEI has 4 diﬀerent defect types. We

divided the two datasets into training and validation sets, as outlined in Table 1 and Table 2, where we also

provide details regarding the original defect classes and their respective total number of instances. In Fig.3, we

illustrate the various defect types in the SEM-ADI dataset. In Fig.4, we present all defect types in the EDR-AEI

dataset. The imbalance in defect classes and the limited availability of data in both datasets prompted us to

develop novel augmentation strategies tailored to SEM and EDR imaging characteristics and assess their impact

on model performance.

3.2 Proposed Data-augmentation strategy

We implemented diﬀerent image augmentation

techniques to enhance the robustness of our model and address

imbalances in the number of instances across defect classes. Within our image augmentation pipeline, utilizing

the Albumentations library,

we integrated the following enhancements to introduce variability in illumination,

noise, and atmospheric conditions:

剩余18页未读，继续阅读

人工智能_SYBH

粉丝: 4w+
资源: 222

SEMI-SuperYOLO-NAS：高NA EUVL半导体缺陷检测的增强方法

半导体缺陷的研究进展

优化YOLOv7用于半导体缺陷检测

基于YOLO系列的目标检测改进算法.docx

用于极紫外光刻非球面光学元件检测的补偿器检测系统

极紫外光刻投影物镜波像差在线检测技术

基于棱镜耦合的无掩模表面等离子体共振干涉光刻的理论分析

Fabrication of a complex-shaped mirror

All-reflective optical system design for extreme ultraviolet lithography

DSP平台下剂量控制板固件研究与开发-论文

vue3-notification:查看3个通知库:speech_balloon:

最新资源