| 摘要: | 二維材料因其在光學和下一代半導體裝置中的廣泛應用而受到廣泛關注。然而，在整合到設備之前，典型的 2D 材料必須從生長基板轉移到目標基板，這個過程通常會引入複雜的缺陷，例如皺紋、破裂和殘留物。確保高品質的石墨烯生產對於這些應用至關重要。 儘管成像技術取得了進步，但高效、準確地檢測轉移的二維材料中的缺陷仍然是一個挑戰。這主要是由於無法精確、快速地辨識這些材料的複雜形態，限制了它們的實際應用。本研究重點關注透過化學氣相沉積 (CVD) 合成的石墨烯，以及隨後使用卷對卷 (RTR)、濕法和乾法轉移技術等方法將其轉移到不同的基材上。引入了創新的自動分割工具，以 JSON（JavaScript 物件表示法）格式產生缺陷註釋，以提高缺陷偵測的一致性和效率。光學顯微鏡 (OM) 影像與 YOLOv7 深度學習模型一起用於準確識別和量化缺陷，特別是那些形狀不規則的缺陷。透過實施精細的資料集分割策略並優化模型的損失函數，偵測精度提高了 10%。此外，自動分割將手動註釋時間減少了 75%，提高了資料一致性，並將預測率提高了 18%。這些發現表明，自動化系統顯著改善了石墨烯的缺陷檢測，簡化了品質控制流程，並提高了生產效率。這些進步預計將支持高性能石墨烯技術的開發，為傳統檢測方法的局限性提供可靠的解決方案。;2D materials have gained significant attention due to their wide applications in optical and next-generation semiconductor devices. However, before being integrated into devices, typical 2D materials must be transferred from the growth substrate to a target substrate, a process that often introduces complex defects such as wrinkles, ruptures, and residues. Ensuring high-quality graphene production is critical for these applications. Despite advancements in imaging technologies, achieving high-efficiency and accurate detection of defects in transferred 2D materials remains a challenge. This is primarily due to the inability to precisely and rapidly recognize the intricate morphology of these materials, limiting their practical application. In this study focus on graphene synthesized via Chemical Vapor Deposition (CVD) and its subsequent transfer to different substrates using methods such as Roll-to-Roll (RTR), wet, and dry transfer techniques. An innovative automated segmentation tool has been introduced, generating defect annotations in JSON (JavaScript Object Notation) format to enhance consistency and efficiency in defect detection. Optical microscopy (OM) images, along with the YOLOv7 deep learning model, were used to accurately identify and quantify defects, particularly those with irregular shapes. By implementing a refined dataset-splitting strategy and optimizing the model′s loss function, a 10% increase in detection accuracy was achieved. Additionally, the automated segmentation reduced manual annotation time by 75%, improving data consistency and increasing the prediction rate by 18%. These findings demonstrate that the automated system significantly improves defect detection in graphene, streamlines quality control processes, and boosts production efficiency. The advancements are expected to support the development of high-performance graphene-based technologies, providing a reliable solution to the limitations of traditional inspection methods. |
