Predicting the Silent Data Error Prone Devices Using Machine Learning

Abstract

Silent Data Errors (SDEs) are a subset of Defective Parts per Million (DPPM) test escapes that cause unnoticed data corruption. Even at very low levels of DPPM, these are visible at cloud service provider data-center scales. In high-volume manufacturing, some defects manifest as SDEs that are screened at system level test (SLT) which is expensive. Due to subtleness of such defects, semiconductor devices prone to SDEs don’t exhibit evident patterns or anomalies in the test data distributions. So, screening such faulty devices with ATE using statistical kill limits is challenging. To accelerate identification of those faulty devices, ahead of system testing, we propose to use Supervised Machine Learning (ML) approach to learn intrinsic patterns in an industrial test dataset. The experimental results illustrate that the embraced supervised learning framework via an ensemble of feature selection methodologies shows a noticeable performance improvement over traditional supervised and unsupervised methods.