AI Vision Transforms Ultrasonic Inspection for Semiconductor Manufacturing
A leading electronics manufacturer was facing a significant quality control bottleneck in their Integrated Circuit (IC) packaging line. The client relies on ultrasonic equipment to inspect ICs for internal defects. However, the machine’s initial First Pass Yield (FPY) was only 50%. To meet production standards, operators had to conduct extensive manual re-inspections to raise the FPY to the target of 95%~98%.
The manufacturer urgently needed an automated, intelligent vision solution to take over this labor-intensive re-inspection process, improve accuracy, and reduce the burden on human operators.
The application inspection challenges included:
Strict Area Thresholds: The inspection criteria were highly specific. A single air bubble could not exceed 3% of the total IC area, and the combined area of all bubbles could not exceed 5% of the total IC area.
Complex Multi-Layer Imaging: The ultrasonic scan evaluates two layers: the first layer for IC bubbles and the second for solder joints.
Interference from “Mapping”: Reflections—or “mapping”—from the soldering layer are often projected onto the IC layer. The system needed the intelligence to recognize these specific reflections and correctly judge them as acceptable products rather than defects.
Irregular Defect Shapes: Air bubbles are naturally random, irregular, and diverse in shape, making it nearly impossible for traditional rule-based machine vision to accurately calculate their areas.
To overcome the inspection hurdles a solution utilizing deep learning to execute advanced image analysis and calculations was implemented:
AI Semantic Segmentation
Instead of traditional machine vision, the system utilizes TM AI Segmentation. This neural network precisely identifies the boundaries of both the chip and the irregular defect areas. By extracting these pixel-perfect shapes, the AI accurately calculates the defect-to-chip area ratio to strictly enforce the 3% and 5% limits.
Seamless Automated Workflow
The ultrasonic equipment outputs the raw images, which are immediately retrieved over the network by TM AI+ AOI Edge. Within the TMflow software, the image contrast is automatically adjusted to make defects pop out, allowing the AI model to make highly accurate judgments in real-time.
Auto-Training and Iteration
The solution incorporated a robust server-based AI Trainer. Operators can review edge cases (such as tricky mapping reflections) and use the Auto Labeling feature to update the database. Through the Auto Training AI system, the factory can automatically collect images, train upgraded models, and seamlessly deploy them back to the edge, continuously improving the system’s intelligence.
System Results & Benefits
High-Precision Quality Control: The AI solution effortlessly automates the complex mathematical task of verifying whether total defect areas fall within the strict <= 5% and <= 3% tolerances, eliminating human subjectivity.
Enhanced Operational Efficiency: By intercepting and processing the 50% of components flagged by the ultrasonic equipment, the AI drastically reduces the sheer volume of manual re-inspections required, saving significant labor costs.
Continuous Improvement: While complex “mapping” reflections are flagged for quick human confirmation during the initial rollout, the closed-loop feedback system ensures the AI model continuously learns from operator inputs, pushing the line toward greater autonomy with every shift.
Transforming Bottlenecked Quality Control Process
The application case study highlights how integrating specialized testing equipment with TM AI Vision can transform a bottlenecked quality control process. By leveraging AI Semantic Segmentation and continuous machine learning, the solution provided a highly accurate, automated, and scalable solution to tackle complex IC inspection challenges.
