本論文是與晶圓廠的產學合作案,該產學案是研究矽晶圓片最上游的製程,又稱為矽晶棒長晶製程。矽晶棒長晶製程對於製程參數與生長環境極為敏感;也就是,一旦參數或環境控制不當,矽晶棒內部原子就會出現堆疊錯位,稱為「斷線」。目前,產學方因長晶製程斷線所造成的成本損失佔所有矽晶圓片製程的一半以上。
目前矽晶圓廠長晶製程是採用人力檢測的方式。然而,人力檢測無法連續且無法準確的進行偵測。基於人力檢測的缺陷,建立長晶製程之即時穩健偵測模型是當急之務。本論文提出了結合即時物件偵測架構及一般化損失函數的長晶製程斷線偵測模型。該模型最佳平均精度均值 (optimal mAP) 達0.99,平均精度均值的平均數 (mAP mean) 達0.94。本論文所提出之即時穩健的斷線偵測模型已證實能大大降低產學方的人力及生產成本。

This research is an industry-academia cooperation project, where the cooperated industry is a well-known silicon wafer fabrication.
This research deals on the single crystal silicon growth process, which is considered as the most upstream process in the cooperative industry.
The single crystal silicon growth process is sensitive to process parameters and growth environment.That is, any minor shift of process parameters or related environment-change will lead to
dislocation of the atom stack, also known as lost structure.
The current frequent lost-structure produces more than 50% of the defective rates among all processes implemented in the above-mentioned cooperative industry.
Motivated by the problem that the current manual lost-structure detection does not permit continuous inspection throughout the process,we investigate and propose an associated real time lost-structure detection method.The proposed method integrates Yolov4 (an object detector) with the optimal parameters via design of experiment and proposed general loss function. The effectiveness of the proposed framework is demonstrated via an estimated optimal mAP 0.99, an estimated mAP mean 0.94. The proposed real-time and robust method helps the cooperative industry reduce huge production and labor cost.

摘要 i
Abstract ii
誌謝 iii
目錄 iv
圖目錄 vi
表目錄 vii
1 緒論 1
1.1 研究背景 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 研究動機與目的 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 符號與名詞定義 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3.1 符號定義 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3.2 名詞定義 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2 文獻回顧 6
2.1 長晶製程晶體缺陷檢測 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 深度學習即時物件偵測方法 . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 YOLO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3.1 YOLO (2016) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3.2 YOLO9000 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3.3 YOLOv3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3.4 YOLOv4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.4 影像前處理方法 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.5 績效指標 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.5.1 混淆矩陣 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.5.2 平均精確度 (Average Precision, AP) . . . . . . . . . . . . . . . . . 12
2.5.3 平均精確度均值 (mean Average Precision, mAP) . . . . . . . . . . 15
3 研究方法 16
3.1 資料標記 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2 深度學習架構 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.3 圖片前處理 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.3.1 Region of interest (ROI) . . . . . . . . . . . . . . . . . . . . . . . . 19
iv
3.3.2 Kindling the Darkness(KinD) . . . . . . . . . . . . . . . . . . . . . 19
3.3.3 局部標準化 (Local Normalization) . . . . . . . . . . . . . . . . . . 20
3.3.4 X 軸標準化 (X-Axis Normalization) . . . . . . . . . . . . . . . . . 20
3.3.5 填入外圍黑色像素值 (Fill black) . . . . . . . . . . . . . . . . . . . . 21
3.4 偵測模型選擇 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.4.1 YOLOv4架構 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.4.2 YOLOv4 損失函數 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.4.3 提出新損失函數 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.5 數據增強 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.5.1 Mosaic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.5.2 Cutmix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.5.3 Mixup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.5.4 Blur . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.6 實驗設計 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.6.1 數據增強方法-部分因子實驗設計 . . . . . . . . . . . . . . . . . 28
3.6.2 超參數選擇-部分因子實驗設計 . . . . . . . . . . . . . . . . . . 29
4 實驗結果 31
4.1 物件偵測模型實驗結果 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.2 數據前處理、數據增強方法實驗結果 . . . . . . . . . . . . . . . . . . . . . . 32
4.3 超參數選擇實驗結果 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.4 研究實驗結果 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.5 偵測模型結合自動化判斷邏輯結果 . . . . . . . . . . . . . . . . . . . . . . . 36
5 結論與未來展望 38
5.1 結論 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5.2 未來展望 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
參考文獻 40

圖目錄
1.1 柴式長晶法生長階段 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 晶棒與晶線示意圖[1] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 長晶製程監控影像 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.1 YOLO (2016) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 confusion-matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 IoU demo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.4 Example:AP calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.5 Example:AP table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.6 Precision-Recall Curve (PRC) . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.7 Adjusted Precision-Recall Curve (PRC) . . . . . . . . . . . . . . . . . . . . 15
3.1 研究方法 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2 長晶監控影像標記 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3 Framework-1 of the Proposed Methodology . . . . . . . . . . . . . . . . . 18
3.4 Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.5 Kindling the Darkness(KinD) structure . . . . . . . . . . . . . . . . . . . . 19
3.6 Step2 result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.7 YOLOv4完整架構 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.8 YOLOv4 Backbone (CSPDarknet53) 架構 . . . . . . . . . . . . . . . . . . . 23
3.9 YOLOv4 Neck, Head 架構 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.10 Mosaic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.11 Cutmix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.12 Mixup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.13 Blur . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.1 三個Policy: Policy 3為本研究方法 . . . . . . . . . . . . . . . . . . . . . . . 36
4.2 自動化斷線偵測機制 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.3 自動化偵測結果 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5.1 Graphical Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

表目錄
1.1 矽晶圓每年需求量 (資料來源: 國際半導體產業協會) . . . . . . . . . . . . 2
2.1 晶體缺陷檢測文獻比較 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 基於MS COCO dataset(test-dev 2017) 之績效比較[12] . . . . . . . . . . . 8
2.3 YOLO 系列模型在COCO 公開資料集之績效表現 . . . . . . . . . . . . . . 10
3.1 部分因子實驗設計 -數據增強方法 . . . . . . . . . . . . . . . . . . . . . 29
3.2 部分因子實驗設計 -超參數選擇 . . . . . . . . . . . . . . . . . . . . . . 30
4.1 YOLO 系列模型運用在本研究資料集之績效表現 . . . . . . . . . . . . . . 31
4.2 模型績效比較 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.3 Effect and P-value . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.4 Optimal design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.5 超參數選擇實驗結果 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.6 Effect and P-value . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.7 研究結果 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

[1] 呂晟瑋. “半導體長晶製程斷線之智能化偵測”. 國立清華大學工業工程與工程管理學系碩士論文 (2020).
[2] Jonathan Tompson et al. “Real-time continuous pose recovery of human hands using convolutional networks”. ACM Transactions on Graphics (ToG) 33.5 (2014),pp. 1–10.
[3] Brian KTanner.Characterization of crystal growth defects by X-raymethods. Vol. 63.Springer Science & Business Media, 2013.
[4] Yongzhao Yao, Yoshihiro Sugawara, and Yukari Ishikawa. “Observation of dislocations in β-Ga2O3 single-crystal substrates by synchrotron X-ray topography,chemical etching, and transmission electron microscopy”. Japanese Journal of Applied Physics 59.4 (2020), p. 045502.
[5] Adeline Lanterne et al.“Characterization of the loss of the dislocation-free growth during Czochralski silicon pulling”. Journal of Crystal Growth 458 (2017), pp. 120–128.
[6] Jun Zhang et al.“A Deep Learning Based Dislocation Detection Method for Cylindrical Crystal Growth Process”. Applied Sciences 10.21 (2020), p. 7799.
[7] Shiori Ueta et al. “Detection of dislocation-free state in Dash-necking process of Si crystal growth furnace using the Czochralski Method”. 2020 59th Annual Conference of the Society of Instrument and Control Engineers of Japan (SICE). IEEE.2020, pp. 695–700.
[8] Yann LeCun et al. “Gradient-based learning applied to document recognition”.Proceedings of the IEEE 86.11 (1998), pp. 2278–2324.
[9] Joseph Redmon et al. “You only look once: Unified, real-time object detection”.Proceedings of the IEEE conference on computer vision and pattern recognition.2016, pp. 779–788.
[10] Jiale Cao et al. “Hierarchical shot detector”. Proceedings of the IEEE International Conference on Computer Vision. 2019, pp. 9705–9714.
[11] Ross Girshick et al. “Rich feature hierarchies for accurate object detection and semantic segmentation”. Proceedings of the IEEE conference on computer vision and pattern recognition. 2014, pp. 580–587.
[12] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. “YOLOv4: Optimal Speed andAccuracy ofObjectDetection”. arXiv preprint arXiv:2004.10934(2020).
[13] Joseph Redmon andAli Farhadi.“YOLO9000: better, faster,stronger”. Proceedings of the IEEE conference on computer vision and pattern recognition. 2017, pp. 7263–7271.
[14] Sergey Ioffe and Christian Szegedy. “Batch normalization: Accelerating deep network training by reducing internal covariate shi”. arXiv preprint arXiv:1502.03167(2015).
[15] Shaoqing Ren et al. “Faster r-cnn: Towards real-time object detection with region proposal networks”. Advances in neural information processing systems 28 (2015),pp. 91–99.
[16] Joseph Redmon and Ali Farhadi. “Yolov3: An incremental improvement”. arXiv preprint arXiv:1804.02767 (2018).
[17] Kaiming He et al. “Deep residual learning for image recognition”. Proceedings of the IEEE conference on computer vision and pattern recognition. 2016, pp. 770–778.
[18] Tsung-Yi Lin et al. “Feature pyramid networks for object detection”. Proceedings of the IEEE conference on computer vision and pattern recognition. 2017, pp. 2117–2125.
[19] Chien-Yao Wang et al. “CSPNet: A new backbone that can enhance learning capability of cnn”. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. 2020, pp. 390–391.
[20] Ali M Reza. “Realization of the contrast limited adaptive histogram equalization(CLAHE) for real-time image enhancement”. Journal of VLSI signal processing systems for signal, image and video technology 38.1 (2004), pp. 35–44.
[21] Karel Zuiderveld. “Contrast limited adaptive histogram equalization”. Graphics gems (1994), pp. 474–485.
[22] Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. “Kindling the darkness: A practical low-light image enhancer”. Proceedings of the 27th ACM International Conference on Multimedia. 2019, pp. 1632–1640.
[23] Diana Sadykova et al. “IN-YOLO: Real-time Detection of Outdoor High Voltage Insulators using UAV Imaging”. IEEE Transactions on Power Delivery (2019).
[24] Tariq M Khan et al.“Efficient hardware implementation strategy forlocal normalization of fingerprint images”. Journal of Real-Time Image Processing (2019), pp. 1–13.
[25] Mark Everingham et al. “ae pascal visual object classes (voc) challenge”. International journal of computer vision 88.2 (2010), pp. 303–338.
[26] Shu Liu et al. “Path aggregation network for instance segmentation”. Proceedings of the IEEE conference on computer vision and pattern recognition. 2018, pp. 8759–8768.
[27] Tsung-Yi Lin et al. “Focal loss for dense object detection”. Proceedings of the IEEE international conference on computer vision. 2017, pp. 2980–2988.
[28] Wheyming Tina Song, Chou Lai, and Yi-Zhu Su. “A Statistical Robust Glaucoma Detection Framework Combining Retinex, CNN, and DOE using Fundus Images”. IEEE Access (2021).