Applications of Machine Learning in the Design of Reliable and Trusted Analog/RF ICs

As electronics continue to penetrate every facet of contemporary life, the analog/RF integrated circuit (IC) market is experiencing unprecedented growth, with its current annual value standing at over $45B. With application domains mainly in wireless communications, real-time control, remote sensing, automotive and health, ensuring reliability and trustworthiness of analog/RF integrated circuits becomes paramount. This tutorial elucidates the role that machine learning and statistical analysis can play towards this end. Specifically, we will discuss (i) classification-based and regression-based test methods for asserting whether the performances of a fabricated analog/RF IC meet its specifications, (ii) statistical calibration methods for tuning the performances of each fabricated device through the use of on-chip knobs in order to increase yield, (iii) spatial and spatiotemporal analysis methods for achieving test time reduction by predicting the performances of a chip based on the performances of other chips in the same wafer or other wafers in the same lot, (iv) statistical side-channel fingerprinting methods for detecting malicious circuit inclusions (a.k.a. hardware Trojans) in wireless cryptographic ICs and for differentiating between genuine and counterfeit chips, (v) the design of on-chip analog neural networks for enabling post-deployment built-in self-test, self-repair and self-trust evaluation, and (vi) statistical methods for attesting the fabrication facility wherein a chip was manufactured. Results from industrial test data and measurements from custom-designed analog/RF ICs will be used to demonstrate effectiveness of machine learning in these applications.