Machine learning opportunities for reliability research

In recent years, Artificial Intelligence (AI) has made impressive progress and many applications are entering the daily lives of everyone. Expectations are that this is merely the beginning of a whole new era in which exciting new applications will arise in fields as for instance medicine, self-driving cars, drug design, robotics, arts, etc. Also in a research environment, machine learning (ML) techniques are being deployed as versatile tools to extract more information from experimental data and from simulation models. In this work, the opportunities offered by machine learning techniques for semiconductor device reliability research are explored. A top-down approach is utilized which starts from existing machine learning algorithms and focusses on opportunities to apply these to enhance the insight in or facilitate the study of degradation physics.

Possible opportunities include (non-exhaustive list):

-use of ML to model simulation results of complex physical situations. This could allow for interpolation in a multidimensional parameter space without having to perform time-consuming additional simulations.

-ML models for classifying and recognizing degradation regimes with complex workloads.

-ML models to understand device-to-device variations, as for example modeling the noise in electrical signals generated by underlying configurations of semiconductor dopant fluctuations or dielectric defects.

Furthermore, it is needed to adapt or (re)design ML models to obtain efficient learning systems. These adaptations include examples like shaping the structure of the model based on physical structural knowledge and properties, selecting physically meaningful features, finding principal components in data, reducing the number of parameters and dimensions to a minimum in order to avoid overfitting, etc. The overall objective of this work is to provide a realistic picture of the gains and limitations of machine learning techniques in reliability research.