Correlation-driven machine learning for accelerated reliability assessment of solder joints in electronics

Vahid Samavatian,Mahmud Fotuhi-Firuzabad,Frede Blaabjerg,Majid Samavatian,Payman Dehghanian

doi:10.1038/s41598-020-71926-7

Abstract

The quantity and variety of parameters involved in the failure evolutions in solder joints under a thermo-mechanical process directs the reliability assessment of electronic devices to be frustratingly slow and expensive. To tackle this challenge, we develop a novel machine learning framework for reliability assessment of solder joints in electronic systems; we propose a correlation-driven neural network model that predicts the useful lifetime based on the materials properties, device configuration, and thermal cycling variations. The results indicate a high accuracy of the prediction model in the shortest possible time. A case study will evaluate the role of solder material and the joint thickness on the reliability of electronic devices; we will illustrate that the thermal cycling variations strongly determine the type of damage evolution, i.e., the creep or fatigue, during the operation. We will also demonstrate how an optimal selection of the solder thickness balances the damage types and considerably improves the useful lifetime. The established framework will set the stage for further exploration of electronic materials processing and offer a potential roadmap for new developments of such materials.

Highlights

The quantity and variety of parameters involved in the failure evolutions in solder joints under a thermo-mechanical process directs the reliability assessment of electronic devices to be frustratingly slow and expensive
While a large number of works has been recently devoted to the use of artificial intelligence in power electronics[30,31,32,33], there have been only a few applied to reliability and lifetime prediction of solder joints in electronic systems
The performance of the regression model in correlation-driven neural network (CDNN) is meaningfully higher than that of the conventional artificial neural network (CANN), indicating that our model includes a set of learning predictive regulation accurately carrying over from one thermal cycling system to another

Summary

Introduction

The quantity and variety of parameters involved in the failure evolutions in solder joints under a thermo-mechanical process directs the reliability assessment of electronic devices to be frustratingly slow and expensive To tackle this challenge, we develop a novel machine learning framework for reliability assessment of solder joints in electronic systems; we propose a correlation-driven neural network model that predicts the useful lifetime based on the materials properties, device configuration, and thermal cycling variations. Mechanics-based acceleration is another approach merging thermal cycling method and mechanical loading to acceptably evaluate the fatigue life of electronic packaging structures[16] With all these descriptions, since it is implausible to enfold all the contributing factors into a single physical framework, a yawning gap between the model outcomes and experimental tests exist, making it a yet to be solved challenge to achieve a high-fidelity thermal fatigue lifetime prediction.

Methods

Results

Conclusion