Machine learning methods for fault classification

Abstract

With the constant evolution and ever-increasing transistor densities in semiconductor technology, error rates are on the rise. Errors that occur on semiconductor chips can be attributed to permanent, transient or intermittent faults. Out of these errors, once permanent errors appear, they do not go away and once intermittent faults appear on chips, the probability that they will occur again is high, making these two types of faults critical. Transient faults occur very rarely, making them non-critical. Incorrect classification during manufacturing tests in case of critical faults, may result in failure of the chip during operational lifetime or decrease in product quality, whereas discarding chips with non-critical faults may result in unnecessary yield loss. Existing mechanisms to distinguish between the fault types are mostly rule-based, and as fault types start manifesting similarly as we move to lower technology nodes, these rules become obsolete over time. Hence, rules need to be updated every time the technology is changed. Machine learning approaches have shown that the uncertainty can be compensated with previous experience. In our case, the ambiguity of classification rules can be compensated by storing past classification decisions and learn from those for accurate classification. This thesis presents an effective solution to the problem of fault classification in VLSI chips using Support Vector Machine (SVM) based machine learning techniques.