Application of Machine Learning in Recognition and Analysis of TSV Extrusion Profiles with Multiple Morphology

In this paper, we propose and utilize a deep convolutional neural network (CNN) model to study the contribution of multiple morphologies and their corresponding mechanisms in through-silicon via (TSV) extrusion. Extrusion surface profiles were obtained from TSV samples with two different pitch distances using white light interferometry (WLI). Visual analysis of a large number of vias revealed three classes of morphologies, i.e. granular, ring, and bulge, that could appear on a single via, individually or simultaneously. In addition, pitch distance was perceived to effectively alter the extrusion morphologies. Leveraging deep learning (DL) approaches, 2408 vias were labeled according to the contribution of each morphology and a CNN-based regression model was developed to efficiently recognize the contribution of these morphologies in TSVs. Achieving a RMSE value of 0.1724, a R2 score of 0.6859, and a top-1 accuracy rating of 82.37%, the proposed model detailed herein demonstrates promising results in recognizing the extrusion morphologies.