Advanced Utilisation of Machine Learning Models for Package Reliability Analysis

   Dr. Karsten Meier 

   Senior Researcher, Assistant Director of Technische Universität Dresden，Germany

Abstract: 

With the progressing employment of heterogeneously integrated packages that come with more complex designs to meet superior performance, miniaturisation as well as limiting energy requirements, development efforts become increasingly challenging. This boosts the interest in virtual development tools finally concluding in the idea of a fully virtual development and qualification scenario. Though, simulation tools that are available as of today often come with their own need for thoroughly development and, at least as important, with solving times in the range of many hours to even a couple of days. With the rise of AI, specifically with the continued efforts to enable the use of machine learning models, neural networkbased models seem to provide a feasible solution for the issue of extensive computing times. The presented work covers how to develop and use neural networks in the context of assessing the thermo-mechanical reliability of electronic packages. In detail, different types of neural networks for a range of purposes will be introduced. Considering the differences between physics-driven and datadriven models, preparation and training requirements will be showcased for the later one. This includes the selection of suitable performance metrics. Finally, the extension of neural network functionality towards the prediction of solder joint stresses under vibration loading at varied temperature conditions will be presented. This will demonstrate the potential of using neural networks in automotive, industrial, avionic and other harsh environment application related reliability assessment. In detail, the prediction of characteristic non-linear plastic strains that can be used for life time predictions will be shown. Furthermore, plastic strain pattern predictions will be demonstrated. Which in fact, extends the assessment of solder joint life time towards the prediction of failure modes.

Speaker's Biography: 

Dr. Karsten Meier is with the Institute of Electronic Packaging Technology at the Technische Universität Dresden (Dresden, Germany) since 2006. After studying electrical engineering he received his Ph.D. from Technische Universität Dresden in 2015. During his studies he spent a research visit at the Packaging Research Center at the Georgia Institute of Technology in Atlanta (Georgia, USA). At the Institute of Electronic Packaging Technology he leads the board level reliability group and is in charge as assistant director. His research activities cover projects on packaging technology developments and package reliability for 5G and automotive applications, power electronics, material characterisation, and thermomechanical simulation incl. FEM and ML which all are source for more than 160 papers he authored or co-authored. Also, he supports a research collaboration with the Center for Advanced Life Cycle Engineering at the University of Maryland (Maryland, USA) on combined mechanical and thermal loadings on solder interconnections and a joint research efforts with the University of Florence on electrical online monitoring for package and board level reliability analyses. He recently chaired the IEEE ECTC sub committee Thermal/Mechanical Simulation and Characterisation and is member of the ECTC executive committee, is a member of the IEEE EPTC sub-committee Advanced Packaging and the IEEE ESTC subcommittee Reliability of Electronic Devices and Systems. As a reviewer he supports the IEEE EuroSimE conference, the ASME and CPMT societies and the Journal of Microelectronics Reliability and other journals.

Speech Outline: 

(1) Motivation to add ML models to the reliability analysis tool set

(2) Brief Introduction of Neural Network Structures

(3) Prerequisites for Creating Data-Driven FNN

(4) FNN Training Approaches and Performance Monitoring

(5) Advancing the Reliability Assessment for Packages under Vibration Loading 

(6) Conclusions 

Who Should Attend: 

The targeted audience includes PhD students, junior but also experienced researchers as well as engineers who work with package and board level reliability simulations incl. Finite Element and Machine Learning modeling with focus on thermo-mechanical load scenarios. This applies for package and board development, understanding of yet existing systems as well as trouble shooting backgrounds. The talk is intended to provide attendees who are interested in gaining knowledge about FE and ML simulations insight, input, or inspiration. Additionally, managers who guide projects or work groups of such scope will learn the major challenges within this field enabling them better communication and supervision skills. The talk will mainly consider Feed Forward Neural Networks to tackle thermo-mechanical reliability analysis but with limited content Recurrent Neural Networks and Long-Short Term Memory, too. Image recognition and related topics will not be covered.