Finite element analysis and neural network model for electronic hidden solder joint geometry prediction

Abstract

This paper investigates an active thermography approach to probing hidden solder joint geometry. Ten boards were fabricated with the same number of solder joints and amount of solder paste (0.061 g), but using three solder joint geometries (60°, 90°, and 120°). The 90° angle solder pin represented a normal joint, and the 60° and 120° angle pins represented abnormal solder joints. Each board was covered with another board that had three openings just big enough to allow the pin terminals to protrude. A semi-automated system was built to heat and then transfer each board set to a chamber where an infrared camera was used to scan the board as it was cooling down. Each board set underwent the heating, cooling, and scanning process for five trials. Two-thirds of the data set was used for model development and one-third for model evaluation. An artificial neural network (ANN) was constructed to predict abnormal joints given thermal data. Results suggest that solder joints with more surface area cool much faster than those with less surface area. A Finite Element Analysis (FEA) of the heating up and cooling down process consistently predicted solder geometry using the ANN with 86% accuracy. This approach can be used not only to inspect bad solder joints (i.e., low reliability) but also to mass screen for cold solder joints during BGA assembly, since the air gaps in cold solder joints may cause them to cool more slowly than normal joints.