Analyzing the Overfitting of Boosted Decision Trees for the Modelling of Stencil Printing

Abstract

Stencil printing is one of the key steps in reflow soldering technology, and by the spread of ultra-fine-pitch components, analysis of this process is essential. The process of stencil printing has been investigated by a machine learning technique utilizing the ensemble method of boosted decision trees. The phenomenon of overfitting, which can alter the prediction error of boosted decision trees has also been analyzed in detail. The training data set was acquired experimentally by performing stencil printing using different printing speeds (from 20 to 120 mm/s) and various types of solder pastes with different particle sizes (particle size range 25–45 µm, 20–38 µm, 15–25 µm) and different stencil aperture sizes, characterized by their area ratio (from 0.35 to 1.7). The overfitting phenomenon was addressed by training by using incomplete data sets, which means that a subset of data corresponding to a particular input parameter value was excluded from the training. Four cases were investigated with incomplete data sets, by excluding the corresponding data subsets for: area ratios of 0.75 and 1.3, and printing speeds of 70 mm/s and 85 mm/s. It was found that the prediction error at input parameter values that have been excluded from the training can be lowered by eliminating the overfitting; though, the decrease in the prediction error depends on the rate of change in the output parameter in the vicinity of the respective input parameter value.