Achieving High Accuracy Pattern Recognition for Die Bonding

When it comes to microelectronics and photonics packaging assembly, components such as die are placed onto packages. All automated die bonding systems require some type of vision processing for the machine to accurately locate the component before picking and placing the part. Because of this dependency on vision processing, the die must possess clear features to reliably assemble packages, especially when it comes to automated processes.

Standard pattern recognition technology, known as auto correlation, uses pixel-to-pixel comparisons. The auto correlation method cannot account for variations such as damaged features and angled presentation. Due to variations of die bond components, it can be difficult to achieve high accuracy and repeatability when it comes to pattern recognition. Scratches, contamination, and even operator handling can cause two components of the same type to look different under the camera.

Another challenge that comes with auto correlation technology is its limited range of adjustable parameters when compared to other methods. For example, in order to raise production quality, one parameter that is typically adjusted is the pattern recognition threshold. By raising this threshold, only higher quality components will be recognized and assembled in production. Although this may seem like an advantage at first, this will lead to a larger amount of rejected components and unnecessary waste.

Furthermore, these challenges can lead to pattern recognition failures, manual references, and operator intervention, all of which are issues that greatly hinder automated production processes. It is critical to accurately find components in order to avoid these problems. Inconsistencies in how a fully-automatic die bonder system finds and locates the learned patterns will result in inconsistencies of die placement. This will not only introduce performance and yield issues, it will also impact downstream processes such as wire bonding.

 

The Solution: VisionPilot®

VisionPilot is a state-of-the-art referencing technology that sets a new standard for vision processing on Palomar’s die bond platforms. In contrast to standard referencing technology, VisionPilot offers a wide range of tools and features to address the typical challenges of pattern recognition. This includes, but is not limited to, radar referencing, synthetic models, and active feedback.

Radar Referencing® is the key feature within VisionPilot that allows the bonder to use shapes instead of standard pixel-to-pixel comparison.

Radar Referencing finds shapes by dividing contrasting shapes Active feedback outlines the part with three colors.

Radar Referencing finds shapes by dividing contrasting shapes.

Active feedback outlines the part with three colors.

 

Synthetic models allow the user to teach perfect shapes as opposed to using a pristine component sample. In this example, due to the manufacturing process of the metallic preform component, there will be a lot of surface variation; no two parts will look identical. By comparing the actual (imperfect) components to a perfect model, the bonder can accurately locate the component.

Active feedback shows the user how accurately the bonder is identifying a component. It uses colors to quickly indicate strengths and weaknesses of that particular reference image.

Download these resources for more information:

Achieving High Accuracy Pattern Recognition for Die Bonding VisionPilot® with Radar Referencing®
Achieving less than 1% voiding in GaN die attach processing VisionPilot Data Sheet

----

Ricky Le
Sr. Applications Engineer
Palomar Technologies, Inc.