In TSMC, machine learning is used to implement auto defect classification (ADC) and keep the accuracy of Advanced Defect Classification (ADC) identification. The machine learning approach we use is human-in-the-loop (HITL), which leverages both human and machine intelligence to create machine learning models. With enough data and human tuning, those machine algorithms can identify and analyze images quickly and incredibly accurate without the need for engineers to teach the machine constantly what exactly defects look like. Therefore, human intelligence will play a significant role in machine learning, which means the experiences and algorithms provided are critical and distinctive for successful machine learning models.
TSMC has widely deployed machine learning in both inline edge computing ADC and offline cloud computing ADC, with a specific focus on key manufacturing stages for tool defect defense. The function and benefit of ADC is to process millions of images daily, improving image defect inspection sampling rate. The inline edge computing ADC, known as “Eagle’s View,” is embedded within the tool to detect defects during processing and isolate affected materials. In contrast, offline cloud computing ADC is inserted post-process to detect defects and hold lots, preventing further loss. The flexibility of machining learning is developed and extended to sustain variant engineering requirements, such as size measurement, defect classification, and particle detection to extend the scale of quality management.
For committed of customer trust, lot-level and wafer-level traceability are now insufficient. Even more, the comprehensive approach to material and die traceability is the reliability requirements for packaging manufacturing.
TSMC has continuously enhanced its data management to achieve full traceability, including per-die mapping back to the original fab wafer position. We utilized 2D barcodes to encode product data and track information such as source wafer positions, bin codes, and engineering experiment labels. Every package receives a unique, individually generated 2D barcode, which is integrated with per-package lithography patterns via a patterning software system. After uploading this unique mark, all process and product information – including process history, tool logs, material data, and production yield – will be listed and searched. Using this “product resume,” we can quickly define the scope of impact for problematic materials or process issues and perform data correlation to analyze the root causes of low yield, thereby minimizing the impact of defects.
