This study aims to develop an empirical model to elucidate the void size formed during an underfilling process of asymmetrical ball grid array flip-chip of I, L and U dispensing orientation with various parameters.
Experimental data of the through-scan acoustic microscope’s dispensing parameters are collected, cleaned and segregated for training and testing purposes. Three machine learning models, support vector machine, random forest and linear regression, are used and evaluated to generate an empirical formulation to explain the correlation between the void percentage and its affecting parameter and tested on samples not used in the model development.
All I-, L- and U-type dispensing orientations point to valve pressure as the leading cause of the voiding phenomenon, with the Naïve Bayes model consistently being one of the better options to perform classification study with these data sets. In addition, the I-type dispensing orientation exhibits the most accurate and consistent regression model, with an error range of 1.43% compared to the other dispensing orientations.
This study provides insight into the voiding phenomenon to dispensing orientations, allowing manufacturers to leverage machine learning to simplify complicated dispensing parameters to void defect size, improving process optimisation and reducing losses due to defective flip-chip underfilling.
The modelling process on the correlation and regression model specific to each I, L and U dispensing method allows a clear insight into the better dispensing orientation to produce a more consistent underfilling region.
