全球低碳發展推動 4H-SiC 高功率電子元件應用，晶體生長與切割過程中所產生的缺陷將影響元件性能，若能提升檢測技術降低瑕疵密度，則可提高量產品質。智慧製造透過提升生產效率、降低成本與提升產品品質，驅動高科技產業變革，現今非破壞性晶圓缺陷檢測主要使用光學檢測方法，本研究檢測系統選用渦電流感測器預計提供更具成本效益且對空間需求較小的檢測方案。由於半導體在常溫常壓下不導電，透過異質結構調整使碳化矽裂紋能夠在渦電流訊號上產生特徵型差異，進而實現渦電流缺陷檢測。
訓練資料集的品質和分佈會影響模型的準確度與泛化能力，本研究選用生成式 AI 進行資料增強以解決訓練數據貧乏的問題，根據感測器收錄數據生成多樣化且逼真的新樣本。缺陷辨識選用半監督式學習方法，先透過非監督式學習判定異常可能值，並將其作為特徵納入監督式學習中，讓模型不僅能準確抓出學習過的缺陷特徵同時擁有新型缺陷識別功能，整合渦電流感測器與半監督生成式 AI 的晶圓缺陷檢測技術架構，建置成本較低且檢測能力穩定的輔助系統，利用智慧製造提升碳化矽在半導體產業的更多可能。

The global push for low-carbon development is driving the application of 4H-SiC high-power electronic components. However, defects generated during crystal growth and cutting processes can affect device performance. Enhancing defect detection technology to reduce defect density can improve product quality in mass production. Smart manufacturing, by increasing production efficiency, reducing costs, and improving product quality, is transforming the high-tech industry. Currently, non-destructive chip defect detection primarily relies on optical inspection methods. This study proposes an eddy current sensor-based detection system, which offers a more cost-effective and space-efficient inspection solution.
Since semiconductors are non-conductive at room temperature and atmospheric pressure, this research employs heterostructure adjustments to induce characteristic differences in eddy current signals caused by SiC cracks, thereby enabling eddy current-based defect detection. The quality and distribution of training datasets significantly impact model accuracy and generalization ability. To address data scarcity, this study utilizes generative AI for data augmentation, generating diverse and realistic new samples based on sensor-collected data. For defect identification, a semi-supervised learning approach is adopted, where unsupervised learning first determines potential anomalies, which are then incorporated as features into supervised learning. This allows the model to accurately detect known defect patterns while also identifying novel defects. By integrating eddy current sensors with semi-supervised generative AI, this study develops a cost-effective and stable defect detection system. This system leverages smart manufacturing to expand the potential applications of SiC in the semiconductor industry.