Modelling the Unmodellable: Algorithmic Fault Diagnosis

In its most basic form, algorithmic fault diagnosis consists of using a fault model to predict the behavior of faulty circuits, comparing these predictions to the actual observed behavior of defective chips, and identifying the predicted behavior(s) which most closely match the observations. The goal of the process is to enable failure analysis by identifying promising locations for further study. The process is successful if the actual defect is contained in the list of possible locations, and if that list is sufficiently small to permit a failure analysis engineer to investigate them. In short, advances continue in both fault models and matching algorithms, which together continue to improve the effectiveness of algorithmic fault diagnosis, in spite of the fact that duplicating the exact behavior of defects remains elusive. A fault model is an abstraction of defective behavior. The most common fault model is the single stuck-at model, which postulates that defects will behave as though a given circuit line was permanently connected to ground (stuck-at 0) or to power (stuck-at 1), and that only a single fault will be present in a circuit at any time. Numerous other fault models have been proposed, such as bridging faults (shorts between lines), open or break faults, IDDQ faults (leakage), delay faults (transition, path delay), etc. In general, the degree of abstraction of a fault model is inversely related to the time needed for dictionary construction. For diagnostic purposes, the capability of the model to directly identify defects is also often (but not always) inversely related to its abstraction level. For example, only a small percentage of defects behave as stuck-at faults, which are relatively abstract, while the less abstract " realistic " bridging faults identified by an inductive fault analysis tool like Carafe [Jee93] or the still less abstract and even more complex faulty behaviors identified by a contamination analysis tool like CODEF [Kha95] will predict an increasing number of behaviors found in actual defective circuits, but at increasing cost. It has been shown in a previous study [Ait95] that a process of fault model validation, using a focused ion beam to insert " defects " into good circuits, can improve the performance of fault models. Ideally, diagnosing faults should be easy. One would simply compare the modeled behavior to the observed and choose the one which matched. Depending on the fault model and the type of defect being diagnosed, this approach …