Combining polynomial regression with unsupervised machine learning on wafer-level packaging reliability prediction

Abstract

ABSTRACT The benefits of wafer-level packaging include better thermal dissipation, lower latency and reduced space consumption. Accelerated thermal cycling test (ATCT) is a regulation that determines whether a product is ready for mass production, but it takes a long time and is costly to perform. The design-on-simulation approach can reduce the number of ATCT experiments and shorten the design cycle. However, the simulation method must be verified before it can be treated as an experiment; if the simulation consistently matches experiments at close range, it can also be treated as an experiment. In addition, the verified simulation method can be used to develop a machine learning (ML) database and obtain an artificial intelligence model for long-term reliability prediction. Due to its effectiveness in solving nonlinear problems with relatively short computation times, polynomial regression (PR) is used in this study as ML model. Results show combining PR with an unsupervised learning algorithm, K-means, can produce more accurate predictions.