Inspecting and Classifying Probe Marks
Deep learning technology helps identify and classify highly variable probe marks to increase efficiency of wafer testing and increase die yields.
Before a wafer is sent to die preparation, all individual integrated circuits are tested for continuity and functional defects. A probe card with dozens of microscopic electrical probes is used for this process. Each probe leaves a small mark on each die as contact is made. This mark should be centered on the die and indicate that the correct amount of pressure was exerted by the probe.
The probe mark is an indicator of accurate prober performance. If the prober works fine, the probe mark shape is good. If the probe is not working properly, the probe mark shape is No Good (NG). For example, if the probe exerts too much pressure, it can become damaged over time and will not perform acceptable electrical testing.
Probes are expensive, so maintaining the correct pressure is important to maintaining their working life. Using traditional rule-based machine vision to detect and classify OK and NG marks is difficult due to the many variations in shape, size, and location of the marks. Inconsistent or false No Good readings can negatively affect yield and chip quality.
Cognex Deep Learning tools make probe mark inspection easier and less time-consuming by helping to verify the difference between OK and NG probe marks. The software is trained from a wide range of images showing correct probe marks and images showing unacceptable probe marks. The unacceptable marks can then be classified as “pressure-related” or “off-center.”
Using this information, operators can adjust the probe pressure or alignment to increase the number of acceptable probe marks and keep the probe in good working order. Using deep learning inspection on probe marks can increase die yields from a wafer compared to alternative methods that can mischaracterize OK marks as unacceptable, or NG marks as acceptable.