Explore a domain-aware hard feature representation with an elastic structure for flip-chip defect detection

The continuous development of flip-chip microelectronics towards ultra-high density and ultra-fine pitch interconnection for aerospace applications has led to the need to detect any manufacturing defects prior to their use. Using a Domain Adaptation (DA) approach, an Elastic space-guided Domain-aware Feature Extractor (EDFE) has been developed to learn and correlate flip-chip vibration signals with defect features. Specifically, EDFE perceives potential features by focusing on the signal frequency response and analyzing the gradient fluctuations of network parameters during DA, exploring unbiased defect features. Then, the elastic equilibrium logic of a spring-mass system is investigated to guide the dynamic evolution direction of network parameters, adjusting the feature density near the decision boundary to enhance the causal relationship between the features and defects. The defect extractor is assessed based on flip-chip vibration signals excited by dual-source air-coupled ultrasonic waves, proving its dual stability of feature learning and detection.