Machine-Learning-Based Error Detection and Design Optimization in Signal Integrity Applications
Roberto Medico, Domenico Spina, Dries Vande Ginste, Dirk Deschrijver, Tom Dhaene · IEEE Transactions on Components Packaging and Manufacturing Technology · 2019
Evaluating the robustness of integrated circuits (ICs) against noise and disturbances is of crucial importance in signal integrity (SI) applications. In this paper, the addressed challenge is to build a software-based framework allowing for automated detection of failures and fast simulation-based evaluation of designs. In particular, these tasks are here addressed using anomaly detection (AD), a branch of machine learning (ML) techniques focused on identifying erroneous or deviant data. In the proposed framework, the ML model only requires the time-domain waveforms and no additional knowledge about the circuit nor about the errors to be identified. Specifically, a two-step approach to detect anomalous behaviors in output waveforms of digital ICs is proposed, comprising a first phase where the ML models are trained to learn relevant features describing the data and a second one where those features are used to identify anomalies with unsupervised or semisupervised AD techniques. Two relevant application examples validate the performance and flexibility of the proposed method.