Anomaly Ranking of LPDDR4 Shmoo Results using Machine Learning

Semiconductors designed for safety-critical automotive applications demand thorough verification prior to mass production. This verification process includes post-silicon characterization of semiconductor devices, such as System-on-Chip (SoC), across the boundaries of their multi-dimensional operating space. Post-silicon characterization is essential for proving interoperability and in-system performance metrics. Unlike production testing, characterization is performed on a smaller sample size, making it critical to identify anomalous SoC behavior within the characterization population, as such anomalies could indicate underlying design defects. The LPDDR4 shmoo test under consideration generates 57,600 shmoo plots, making manual analysis impractical due to volume of data, the subtlety of potential anomalies, and the time-intensive nature of the task. To address this challenge, this paper proposes a novel machine-learning-based method that employs a reconstruction-based anomaly detection algorithm, specifically an autoencoder, to rank anomalous behaviour in the shmoo test results. The results of the proposed approach show that all anomalous results are ranked within the top 5% of the 57,600 shmoo test results. By focusing only on the top-ranked results, the proposed method reduces the manual effort by 95%, enabling more accurate and repeatable anomaly detection.