Exploring the Data-Driven Modeling Methods for Electrochemical Migration Failure of Printed Circuit Board

Abstract

Driven by the integration development of electronic products, the wiring density on printed circuit board (PCB) becomes higher continuously, which causes the possibility of electrochemical migration (ECM) of PCB under application environment is increased correspondingly. Although the life models of ECM has been developed on single or multiple influencing factors, such as the temperature, the relative humidity, and the electric field strength on PCB, by empirical fitting or failure physics, the more complicated environmental conditions, such as dust contamination, made it further difficult for life modeling of ECM. Therefore, data-driven regression methods were proposed to realize the life modeling of ECM on PCB. In this paper, three methods of machine learning were applied on the life modeling of ECM, involving in support vector regression (SVR), gradient boosting regression trees (GBRTs), and random forests (RFs) regression algorithms. Taken as a representative PCB, the standard comb-pattern PCB was carried out the temperature humidity bias (THB) test to simulate the ECM process. The time to failure (TTF) of PCB was obtained from the surface insulation resistance curves during the simulation tests and forms the training and test set. The normalized mean squared error (NMSE) was used to explore the optimal parameters of these algorithms and to evaluate the life prediction effect of ECM on PCB. The modeling methods by machine learning were analyzed and compared with the model based on failure physics to prove the validity even under limited data size.