Neural networks for enhanced stress prognostics for encapsulated electronic packages - A comparison

https://doi.org/10.1016/j.microrel.2021.114181Get rights and content

Highlights

  • The stress distribution in encapsulated standard packages is examined using different types of neural networks.

  • Measurement data is interpreted as time series data and used to predict the development of the stress.

  • Gated recurrent neural networks provide the most promising prognostic results.

  • This work is meant to contribute to establish data driven models to be compared and eventually merged with PoF based models.

Abstract

The prediction of high-resolution mechanical stress distributions in electronic chips with a view to improving prognostic and health management in electronics and N/MEMS via artificial intelligence-based processing of measurement data is the focus of this study. Temperature, shear, and differential stress time-series data acquired through piezo-resistive silicon-based stress sensors on multiple electronic packages inside a thermal shock chamber were monitored, recorded, and subsequently analyzed by various neural network models to create a better understanding of the failure behavior over time including failure mechanisms, the delamination process in particular and the related stress distribution. Monitoring stress changes via continuous observation of material stiffness and interface integrity as factors influencing the local boundary conditions on chip cells, conveys pivotal information concerning the degradation progression. Deep neural networks empowered by backpropagation were trained to predict the stress distributions and ultimately monitor the degradation based on time-series data and were subsequently assessed for their performance to reliably predict in-plane stress developments and distributions on chips. In this study, long-short-term-memory- and gated-recurrent-unit-based networks could accurately predict the behavior of a single chip with smallest error.

Introduction

With the advent of artificial intelligence (AI), learning algorithms are becoming progressively more efficient in failure root-cause analysis and subsequently at anticipating the likelihood of failures and at prognostics and health management (PHM) by collecting reams of performance data for microsystems and microelectronic packages. So AI has the potential to handle lifetime prediction based on big data analysis and thus contribute substantially to certain reliability challenges, as for example future safety paradigms for autonomous driving. In this paper, deep learning (DL) methods have been utilized as pivotal tools to identify the complex relationship between independent and interdependent variables and features, as it focuses on forecasting high-resolution stress distributions through time-series data on industry-grade samples evaluated during stress testing. This work is meant to contribute to establish data driven neural network models to be compared and eventually merged with physics-of-failure (PoF) based finite element models. Data driven models do not focus on a specific local failure parameter but rather use one or more potential failure indicators by condition monitoring to predict a critical situation. This work addresses the first paradigm, whereas a physics-of-failure approach and detailed stress-versus-delamination correlations will be sought after in another paper.

Currently, there is a paucity of research into functional data analysis algorithms in order to develop reliable and sophisticated predictive models for reliability assessments of electronic packages, which may also be caused by a lack of meaningful data. However, in long term, the AI-based approach may facilitate the lifetime approximation of electronic packages. At present, the purpose of this study is the use of different deep neural network (DNN) models to create and benchmark a robust prediction, which could be used for reliability assessments in heterogeneously integrated microelectronic packages [1].

A common failure mode observed in electronic packages is the delamination at different sections of the package, which starts with a crack formation and continues growing at the interface of two dissimilar materials (cf. Fig. 1). The most common cause for such behavior is the internal stress existing at the interfaces, usually caused by the difference in the coefficient of thermal expansion and/or moisture preceded by interfacial defects, see e.g. [2,3] for reference. The physics-of-failure based lifetime prediction of interface delamination by fracture mechanical methods is subject of current R&D work [[4], [5], [6]], which represents the flipside of AI-based methods to be merged into hybrid models for future functional safety concepts [7].

In a previous study [8], test vehicles (TV) were used to create a prognostic model and neural networks (NN) were trained to predict the degradation of the packages. The trained NNs were capable to retrieve information about the structural changes in the packages, including delamination. The final delamination is monitored using scanning acoustic microscopy (SAM) as shown in Fig. 2.

Section snippets

Introduction on artificial intelligence, machine learning, and deep learning

Artificial intelligence (AI) technology matured over the last few years into a disruptive technology. Breakthroughs in a technique called deep learning set new performance levels for various AI applications. The currently evolving fields include predictive maintenance [9,10], quality control [11], smart automation [12] and process optimization [13].

Quite often, the terms machine learning (ML) and AI have been used interchangeably, which is not true. ML is a part that helps the AI to learn and

Experimental background

In order to improve the PHM in electronic packages, one option is to acquire useful data relevant to their reliability, called failure indicators. In this research, mechanical stress distributions were used to gather this type of data. Throughout the data acquisition process, stress measurements were collected during experiments from sensors (cf. Fig. 4, Fig. 5), embedded in test vehicles (TVs). The data was recorded in a structured way.

The TVs, the setup and the method of data collection is

Overview on neural networks

In this study, six prognostic configurations were trained to predict the stress distributions based on time-series data. In all cases, the training is empowered by backpropagation of error [22] and therefore all used algorithms originate from the class of supervised learning algorithms. During the training phase, input-output pairs are used to compute the loss function, describing the overall error of the prediction, and the gradient of the loss with respect to the given weights w used by the

Stress prediction using neural networks: results, discussion and comparison

During the training phase, all six networks have been trained using the same scheme. Following a 95% vs. 5% split, the available data (cf. Section 3, 4450 and 4444 measurements points, respectively) for each chip was divided into a training set and a much smaller validation set. The networks are trained multiple times, given by the number of epochs, on the training data. In the here presented case, the number of epochs is given by 50.

Conclusions and outlook

Results of the current study have shown, that gated-RNN models such as LSTM and GRU networks can provide promising prognostic results on stress distribution inside of encapsulated standard packages in different use cases. The simpler networks studied in the paper are generally quicker trained than the gated-RNN models but are unable to successfully fulfill the given tasks.

In this research, it was shown that certain neural networks can accurately predict the behavior of a single chip; however,

CRediT authorship contribution statement

  • Peter Meszmer: Conceptualization, Methodology, Software, Validation, Writing, Visualization, Supervision, Project administration

  • Mehryar Majd: Methodology, Software, Formal analysis, Investigation, Data curation, Visualization

  • Alexandru Prisacaru: Investigation, Resources

  • Przemyslaw Jakub Gromala: Validation, Supervision

  • Bernhard Wunderle: Conceptualization, Validation, Supervision.

Declaration of competing interest

None.

References (41)

  • A. Prisacaru et al.

    Prognostics and health monitoring of electronic system: a review

  • B. Wunderle

    Thermo-mechanical reliability during technology development of power chip-on-board assemblies with encapsulation

    Microsyst. Technol.

    (2009)
  • B. Wunderle et al.

    Lifetime modeling for microsystems integration - from nano to systems

    J. Microsyst. Technol.

    (2009)
  • E.A. Poshtan et al.

    An in-situ numerical-experimental approach for fatigue delamination characterization in microelectronic packages

  • Schlottig, G., Pape, H., Wunderle, B. & Ernst, L. J. Establishing the critical fracture properties of the die...
  • Hager, M., Gromala, P., Wunderle, B. & Rzepka, S. Affordable and safe high performance vehicle computers with...
  • A. Prisacaru

    Degradation prediction of electronic packages using machine learning

  • C. Escobar et al.

    Machine learning techniques for quality control in high conformance manufacturing environment

    Adv. Mech. Eng.

    (2018)
  • A.C. Müller et al.

    Introduction to Machine Learning With Python and Scikit-Learn

    (2015)
  • D. Cournapeau

    Scikit-learn

  • Cited by (5)

    1

    Former student.

    View full text