Authored By:
Thaer Alghoul, Pubudu Goonetilleke, and Chris Alvarez
Intel Corporation
Hillsboro, OR
Summary
Solder joint reliability is determined by multiple factors, one of which is voiding. The formation of voids in solder joints is caused by entrapped air during the reflow process and could be challenging to eliminate entirely. When voiding exceeds a certain level, it may lead to joint failure and is therefore important to quantify. X-ray inspection is a nondestructive method that can be used to measure voiding, but currently available X-ray equipment has limitations. Automated X-ray Inspection tools (AXI) are fast but lack accuracy, whereas 2D X-ray tools are accurate but slow and cannot be used in aproduction environment.
We will be showing a new methodthat we developed using Deep Learning (DL) to improve thespeed of void measurement with a 2D X-ray tool while stillmaintaining accuracy. Our DL method is a two-phaseapproach. The initial phase involves the detection of solderball area and then segmentation to detect the boundaries ofthe solder ball. The second phase involves segmentation todetect voids. We have achieved 99.9% solder ball detectionand area segmentation, and 99.5% void segmentation. Thecapability of the Deep Learning method used is thendetermined using Measurement Capability analysis.
Conclusions
Solder joint voiding and width measurements collection is an important metrology to assess the reliability of solder joints and to predict possible defects in BGAs. Solder joint mathematical modelling is a driver to improve data collection methods used to compare real life data with predicted data. In order to collect real life data, we had to collect a sufficient number of accurate measurements in a timely manner. Collecting measurements from 2D X-ray images of solder balls and processing them using AI helped in reducing the time to create SJWVM recipes as well as reducing operators’ effort to collect and process solder ball data.
A measurement capability analysis was performed to certify the new process and determine its capability to produce reliable measurements that can be used in SJMM. The MCA showed the data collected had an accuracy with ±5μm for width measurements and ±10% for voiding measurements. Results using our new process showed a reduction in data collection time and an increase in collected data.
Initially Published in the SMTA Proceedings
|
