Mittal_cmu_0041E_10522.pdf (5.65 MB)

Learning Enhanced Diagnosis of Logic Circuit Failures

Download (5.65 MB)
posted on 13.03.2020 by Soumya Mittal
As semiconductor manufacturing progresses to smaller process nodes, it is becoming increasingly difficult to climb the yield learning curve rapidly. The rate of yield learning dictates the growth and success of the semiconductor industry, and must be accelerated to fulfill competitive time-to-market, time-to-money and time-to-volume requirements. Software-based diagnosis plays a crucial role in yield learning. Diagnosis comprehends the test response of a failing circuit to determine the location, and sometimes, in addition, characterize the nature of a defect affecting the failing circuit. Besides identifying likely failure mechanisms and increasing the quality of chip testing, the feedback provided by diagnosis is used to select chips for physical failure analysis (PFA). PFA aims to visually examine a chip to characterize a defect, prevent similar defects in the future and, consequently, improve the design and manufacturing of a chip.However, PFA is often destructive, time- and cost-intensive, and not always successful. Diagnosis, on the other hand, is non-invasive and time- and cost-effective; moreover, it assists PFA and guides yield learning. The advantages of diagnosis, coupled with the diminishing performance of PFA with advancing technology, make it an encouraging facilitator for rapid yield learning.Therefore, the objective of diagnosis must be logic-level defect characterization to minimize (and ideally eliminate) the need for PFA, and accelerate yield learning. Logic characterization of a defect includes the derivation of its physical location and precise logic behavior. In this dissertation, a comprehensive diagnosis methodology is developed to actualize the aforementioned objective.The developed methodology comprises of three methods. LearnX/MD-LearnX is a physically-aware method that employs (a) the X-fault model to avert the elimination of a correct defect candidate and (b) machine learning to build a candidate-ranking model that learns the hidden correlations between the tester response and the defect candidates to pinpoint the correct candidate.
PADLOC, which stands for Physically-Aware Defect LOcalization and Characterization, improves the physical location of a back-end defect (i.e., a defect that affects one or more interconnects and resides outside a standard cell) returned by LearnX/MD-LearnX by partitioning the defective net into physical subnets and identifying the subnets that influence defect excitation. In addition, PADLOC deduces the precise impact of a defect on the circuit functionality by examining its surrounding circuitry.
NOIDA, which stands for NOise-resistant Intra-cell Diagnosis Approach, pinpoints the location of a defect within a failing standard cell implicated by LearnX/MD-LearnX. In contrast to prior work that typically constructs/employs a fault dictionary, NOIDA ascertains the location as well as the behavior of a front-end defect (i.e., a defect that resides inside a standard cell) by monitoring the logical activity of its intra-cell neighborhood. Additionally, NOIDA is resistant to circuit-level noise that may originate from potentially inaccurate transistor-level simulation.
Results from numerous experiments reveal that our diagnosis methodology outperforms state-of-the-art commercial diagnosis. LearnX/MD-LearnX reports fewer defect candidates than commercial diagnosis for 69.4% silicon fail logs without losing accuracy. PADLOC implicates a smaller physical area for a defect for 47.2% silicon fail logs and attains at most 44X improvement. NOIDA reports an ideal diagnosis for 38.0% more front-end defects, when compared to leading-edge commercial diagnosis. In the presence of noise, NOIDA achieves an ideal diagnosis 7.6X more often.
In summary, this dissertation endeavors to characterize a defect residing in a logic chip in terms of its precise physical location and logic behavior, which, consequently, most likely, enables rapid yield learning. The deployment of machine learning to pinpoint the correct candidate in LearnX/MD-LearnX, and the investigation of the neighborhood of a defect to determine its exact physical location and logic behavior in PADLOC and NOIDA are the novel components of this dissertation, and the reasons for its superiority over the state-of-the-art.




Degree Type



Electrical and Computer Engineering

Degree Name

  • Doctor of Philosophy (PhD)


Shawn Blanton