AI Semiconductor Quality Control

In semiconductor manufacturing, almost no defect is acceptable: one contaminated part can fail in a customer's product and lead to recalls. Manual inspection is slow and uneven, inspectors get tired, and defects measured in micrometres are easy to miss at line speed.

Golden Sea built a vision inspection system that sits on the production line. Industrial cameras photograph each part as it passes the station. A YOLOv8 model, fine-tuned on thousands of labelled defects supplied by the client, sorts each part in under 50 ms: pass, or one of more than 15 defect types such as cracks, chipped edges, particles, misalignment and solder bridges.

Real defects are rare, so the training data combined real samples with synthetic ones: rotated, rescaled, relit, and defects blended onto clean parts. That gave the model enough examples of each defect to learn from.

A defect can stop the line automatically and send the part for human review or rejection. Every inspection is logged with the part ID, the image, the defect type, the confidence and what happened to the part. Shift reports show defect trends and warn QC managers when one type starts rising, which often means a setting upstream is drifting.

In production the system reached 99% defect detection accuracy and raised QC throughput by 25 to 40%.