隨著半導體製程邁入極致微縮的2奈米節點，傳統的自動光學檢測（AOI）正面臨物理極限與高誤報率（Overkill）的挑戰。為了突破檢測瓶頸、降低不良品質成本並維持高良率，導入人工智慧（AI）已成為全球半導體企業競爭力的關鍵戰略，並使品質管理成為策略武器。本研究探討全球晶圓代工龍頭T公司應用AI技術持續精進晶圓缺陷檢測的技術發展、管理挑戰及未來趨勢。
透過個案研究法，蒐集次級資料並輔以專家深度訪談，運用PEST、SWOT、內容分析法等綜合分析。發現T公司將缺陷檢測由傳統的規則導向（Rule-based）成功轉化為數據導向（Data-driven）的AI技術架構（如卷積神經網路CNN），提升缺陷分類準確率至 99.2% 以上，顯著降低誤報率27%，並減少人為漏判率40%至60%。將 AI 邊緣運算檢測與先進製程控制（APC）及生產履歷追蹤系統深度結合，即時進行真因分析（RCA）與機台參數動態補償，大幅加速良率學習曲線10%至15%。精準的良率控制從源頭減少了水電與化學品的報廢浪費，創造極佳的綠色製造與 ESG 協同效應。在技術與組織的挑戰上，克服資料不平衡、邊緣運算硬體限制、資訊安全防護，以及跨領域人才的溝通障礙。另外，多模態（Multimodal）感測融合、生成式AI（如 GANs）擴增少數缺陷樣本與邊緣AI閉環自適應控制是未來智慧工廠的趨勢。最後，本研究建議將 AI 檢測視為企業核心數位資產，導入聯邦學習以兼顧供應鏈數據合作與營業秘密保護，輔以綠色AI投資評估，建構長期競爭護城河。

As semiconductor manufacturing advances toward the extreme scaling of the 2-nanometer node, traditional Automated Optical Inspection (AOI) faces severe physical limitations and high overkill (e.g. false positive) rates. To overcome these inspection bottlenecks, reduce the cost of poor quality, and sustain high yields, integrating Artificial Intelligence (AI) has become a crucial competitive strategy, making quality management a strategic weapon. This study explores how Company T, a leading global wafer foundry, applies AI to continuously refine defect inspection, examining its technological development, managerial challenges, and future trends. Using a case study approach, secondary data and in-depth interviews are evaluated through PEST, SWOT and content analyses. It is found out that Company T successfully transitioned from traditional rulebased methods to data-driven AI architectures (e.g., Convolutional Neural Networks), elevating defect classification accuracy above 99.2%. This transformation significantly reduced overkill rates by 27% and human underkill rates by 40% to 60%. By deeply integrating AI edge computing with Advanced Process Control (APC) and production tracking, the company achieves real-time Root Cause Analysis (RCA) and dynamic parameter compensation, accelerating the yield learning curve by 10% to 15%. This precise yield control minimizes water, power, and chemical waste at the source, generating excellent ESG synergies. Organizationally and technically, the company navigated challenges regarding data imbalance, edge hardware constraints, cybersecurity, and cross-disciplinary communication. Future smart factory trends highlight multi-modal sensor fusion, Generative AI (e.g., GANs) for rare defect augmentation, and edge AI closed-loop adaptive control. Ultimately, the study suggests treating AI inspection as a core digital asset, adopting federated learning for secure supply chain collaboration, and utilizing green AI evaluations to construct a long-term competitive moat.

一、中文文獻
T公司113年度永續報告書 (2024)。ESG年執行成果。檢自: https://esg.tsmc.com/zh-Hant/file/public/2024-TSMC-Sustainability-Report-c.pdf。
T公司113年度永續報告書 (2024)。數位卓越競賽帶動AI創新與環境友善雙贏。檢自: https://esg.tsmc.com/zh-Hant/file/public/2024-TSMC-Sustainability-Report-c.pdf。
T公司113年度年報 (2024)。T公司財務、營運與永續成就。檢自: https://investor.tsmc.com/sites/ir/annual-report/2024/2024-Annual%20Report-C.pdf。
T公司113年度年報 (2024)。2.2.3 產業未來展望、機會與挑戰 -
車用電子。檢自: https://investor.tsmc.com/sites/ir/annual-report/2024/2024-Annual%20Report-C.pdf。
T公司官方網站 (2026)。Automation in Packaging Fab。檢自: https://www.tsmc.com/chinese/dedicatedFoundry/services/apm_intelligent_packaging_fab/intelligentFab_automation。
Wu, S. (2025)。良率維持95% 關鍵：AI 導入半導體製程案例。檢自: https://www.powerarena.com/zh-hant/blog/yield-95-ai-in-semiconductor-manufacturing/
呂建興 (2025)。智慧檢測賦能異質整合：AI驅動的半導體先進製程缺陷檢測技術發展。機械工業雜誌，511，6-10。https://doi.org/10.30256/JIM.202510_(511).0004。
林沁瑀（2025）。基於深度學習之半導體晶圓缺陷分析決策優化方法之研究 (碩士論文)。私立中原大學，臺灣桃園市。
品質協會（2017）。COPQ是什么意思？不良品质成本(Cost of Poor Quality)的分类和作用。檢自: https://www.pinzhi.org/thread-566-1-1.html。
施宏儒 (2025)。應用生成式對抗網路於碳化矽晶圓表面缺陷資料增強與檢測之研究 (碩士論文)。國立勤益科技大學，臺灣台中市。
洪宗貝、李詠騏、蘇子仁 (2023)。整合深度學習與自動化技術於半導體瑕疵檢測之解決策略。第二十八屆人工智慧與應用研討會 (TAAI 2023)。檢自: https://www.taai.org.tw/files/domestic/139/TAAI-camera-ready-139.pdf。
財訊新聞中心（2026）。2025Q4全球前10大晶圓代工產值季增2.6%。檢自: https://www.wealth.com.tw/articles/e31408df-37cc-4e8e-9755-fbe498929514。
張啟敏（2026）。AI算力95%押在T公司？分析師示警「單點風險」成隱憂。檢自: https://www.storm.mg/article/11106387。
黃仲宏（2024）。半導體工業中的自動光學檢測發展。機械工業雜誌，499，6–11。https://doi.org/10.30256/JIM.202410_(499).0004
鄭淵引（2022）。基於生成對抗網路之資料擴增並應用於晶圓瑕疵分類（碩士論文）。國立陽明交通大學，臺灣新竹市。
鄭景尤（2014）。電子束吞吐量偏低 光學檢測仍居先進製程主流。新電子科技雜誌。檢自: https://www.mem.com.tw/%E9%9B%BB%E5%AD%90%E6%9D%9F%E5%90%9E%E5%90%90%E9%87%8F%E5%81%8F%E4%BD%8E%E3%80%80%E5%85%89%E5%AD%B8%E6%AA%A2%E6%B8%AC%E4%BB%8D%E5%B1%85%E5%85%88%E9%80%B2%E8%A3%BD%E7%A8%8B%E4%B8%BB%E6%B5%81/。
魏鑫陽（2025）。IDM2.0是什麼？為何Intel陳立武可能力推求翻身？。檢自: https://www.gvm.com.tw/article/120209

二、英文文獻
Chien, J.-C., Wu, M.-T. & Lee, J.-D. (2020). Inspection and Classification of Semiconductor Wafer Surface Defects Using CNN Deep Learning Networks. Applied Sciences, 10(15), 5340. https://doi.org/10.3390/app10155340.
Conway, D. (2010). The data science Venn diagram. Drew Conway. Retrieved from: http://drewconway.com/zia/2013/3/26/the-data-science-venn-diagram.
Pang, G., Shen, C., Cao, L., & Van Den Hengel, A. (2021). Deep learning for anomaly detection: A review. ACM Computing Surveys (CSUR), 54(2), 1-38. https://dl.acm.org/doi/abs/10.1145/3439950.
Patsnap Eureka (2025). How T公司 Implements AI-Based Defect Inspection at 3nm. Retrieved from: https://eureka.patsnap.com/article/how-tsmc-implements-ai-based-defect-inspection-at-3nm.
Peters, L. (2026). Catching Critical Defects In TSVs And Stacked Chips. Semiconductor Engineering. Retrieved from: https://semiengineering.com/catching-critical-defects-in-tsvs-and-stacked-chips/.
Robovision. (2025). Inspection Beyond Rules: How Vision AI Accelerates Yield Ramp for OEMs. Retrieved from: https://robovision.ai/blog/inspection-beyond-rules-how-vision-ai-accelerates-yield-ramp-for-oems.
Sverdlik, D. (2025). AI quality control in manufacturing: Reducing errors across 5 critical workflows. Retrieved from: https://xenoss.io/blog/ai-manufacturing-quality-control.
Wang, C.-C., & Chien, C.-H. (2025). Machine Learning for Industrial Optimization and Predictive Control: A Patent-Based Perspective with a Focus on Taiwan’s High-Tech Manufacturing. Processes, 13(7), 2256. https://doi.org/10.3390/pr13072256.
Yang, Y.-F. (2019). A deep learning model for identification of defect patterns in semiconductor wafer map. 2019 30th Annual SEMI Advanced Semiconductor Manufacturing Conference (ASMC). Retrieved from: https://doi.org/10.1109/ASMC.2019.8791815.