Predicting the Transfer Efficiency of Stencil Printing by Machine Learning Technique

Experiment was carried out for acquiring data regarding the transfer efficiency of stencil printing, and a machine learning-based technique (artificial neural network) was trained for predicting that parameter. The input parameters space in the experiment included the printing speed at five different levels (between 20 and120 mm/s) and the area ratio of stencil apertures from 0.34 tol.69. Three types of lead-free solder paste were also investigated as follows: Type-3 (particle size range is 20–45$\mu$m), Type-4 (20–38$\mu$m), Type-5 (10–25$\mu$m). The output parameter space included the height and the area of the print deposits and the respective transfer efficiency, which is the ratio of the deposited paste volume to the aperture volume. Finally, an artificial neural network was trained with the empirical data using the Levenberg – Marquardt training algorithm. The optimal tuning factor for the fine-tuning of the network size was found to be approximately 9, resulting in a hidden neuron number of 160. The trained network was able to predict the output parameters with a mean average percentage error (MAPE) lower than 3 %. Though, the prediction error depended on the values of the input parameters, which is elaborated in the paper in details. The research proved the applicability of machine learning techniques in the yield prediction of the process of stencil printing.