Machine Learning Techniques for Testing and Diagnosis

Manuel Barragan  , Université Grenoble Alpes, CNRS, Grenoble-INP, TIMA, FR


In the last few years, a wide variety of machine learning-based strategies have been proposed for simplifying and enhancing standard test and diagnosis procedures. In essence, these indirect techniques replace the direct measurement of functional specifications by simpler signatures that are strongly correlated to the target results. Advanced machine learning models are employed to map the signatures to the target variables. In this seminar we will explore a variety of these machine learning-based solutions in different application scenarios. The seminar will follow a tutorial approach with practical case studies in order to highlight the different stages in the definition of a machine learning-based test protocol, the potential pitfalls at each stage and recommended best practices.


The lecture will cover the following topics:

  • Machine learning for indirect test, diagnosis and calibration of integrated circuits
  • Yield learning, yield enhancement
  • Regression models and training
  • Feature selection, extraction and design
  • Pre-silicon validation: Extreme Value Theory; Statistical Blockade; Modified Monte Carlo sampling techniques
  • Limitations and best practices: Dealing with defects; the defect filter and outlier screening; correlation vs causation; Markov’s Blanket as a defect filter; Questioning Monte Carlo validation; Bayesian Model Fusion; comments on complex calibrated AMS systems