Classification of Microelectronics Radiation Effects Using Unsupervised Machine Learning

This paper describes an unsupervised machine-learning method for detecting single-event transient (SET) radiation effects in microelectronics and classifying areas within circuits sensitive to radiation effects. A neural network creates latent spaces for classifying transient phenomena and minimizes loss through the mean squared error function. Differences in the loss within the latent space allow the categorizing of major groups of unique SET responses. A data set including 100,000 transient waveforms obtained through carrier generation based on sub-bandgap two-photon absorption (TPA) laser-induced single event effects is used to demonstrate the data clustering technique. After initial model training, the testing data set is placed in the latent space and used to train a K-Means model, which allows for automatic clustering of related waveforms. The clustering technique is validated by illustrating the data clusters on a two-dimensional spatial mapping of sensitive regions obtained from the TPA experiments. The method can quickly identify radiation-induced anomalies, providing near real-time assessment of measured data. It provides a computationally and temporally cheap way to analyze and group SET anomalies without examining the entire data set or using a supervised technique requiring labeled training data. The model can ingest any new incoming data and predict the disruption's originating location.