中文 English

Breakthrough For Scan Diagnosis With Machine Learning

Improve yield and failure analysis by detecting, refining, clarifying, and resolving defects inside standard cells.

popularity

Cell-aware diagnosis is a new and effective way to detect defects inside standard cells. Industry standard failure analysis (FA) results from a major foundry show that cell-aware diagnosis is very effective at increasing the resolution of the diagnosis by reducing the number of suspects in cell-internal defect data.

With advanced technology nodes, we have more complex layout structures and fabrication processes, like complex finFETs and multi-patterning. One side-effect is that more defects and systematic yield issues are expected in the front-end layers that are inside library cells. Defective die are subjected to FA to find the root cause of yield loss, but the high cost and low success rate of FA on finFETs demands new solutions. When yield engineers have to deploy nanoprobing and TEM imaging, they want to have more confidence that those expensive techniques will be successful in finding defects that can lead to yield improvement­. Therefore, it’s more important than ever to carefully select failing die that are most likely to represent main yield loss mechanisms. Example defects are shown in Figure 1.


Figure 1. A layout of an OR cell showing two example defects; a Metal1 short and a poly open. The defect cell models are shown on the right.

For digital logic, you could build a defect mechanism Pareto for a volume of failing die based on their scan test diagnosis results, but the result will include fake suspects. How do you cut through the ambiguity, or diagnosis noise?

After years of research, and in collaboration with fabless semiconductor manufacturers, foundries, and integrated device manufacturers, Mentor developed root-cause deconvolution (RCD), an unsupervised machine learning algorithm to estimate the defect Pareto from volume diagnosis results in the presence of noise. RCD has successfully reduced overall cost and cycle time to finding the root cause of back-end yield loss in the interconnect.

Now cell-aware volume diagnosis, with the machine learning RCD algorithm, is extended to front-end defect mechanisms inside library cells in the advanced nodes. The new cell-aware diagnosis technology allows diagnosis to report suspects inside the cells. It creates a defect Pareto that includes both back-end interconnect and front-end cell-internal physical defect mechanisms.

Cell-aware diagnosis leverages fault models derived from analog simulation and uses a failure data collection and diagnosis flow identical to that of traditional diagnosis. The RCD algorithm estimates the probability distribution of a defect mechanism in a population of failing die by building a statistical model, then calculating unknown root cause distribution using maximum likelihood estimation. Basically, the distribution that maximizes the likelihood of seeing the given set of diagnosis reports must be the correct one. Once the root cause distribution is calculated, the probability that a suspect is the real defect in a failing die can be determined.

GlobalFoundries conducted a controlled injection experiment to test the cell-aware diagnosis methodology. A population of 200 die was created by injecting polysilicon bridges in a design. First, the failures were diagnosed with non-cell-aware diagnosis and also run through the RCD algorithm. The results place several big cells and a metal 1 open at the top root causes in the Pareto chart. Useful, but not very precise.

Next, the same die were processed with cell-aware diagnosis and run through RCD. The Pareto chart then included the exact target root cause, i.e., polysilicon shorts, and the fake top root cause of metal 1 open was eliminated. This highlights the tremendous value of cell-aware diagnosis with RCD in resolving physical root causes in the front-end layers inside the cells.

With such precise root cause distribution, die selection for FA can be more precisely targeted, resulting in a significant reduction in overall yield analysis cycle time and cost.

The experiment was repeated on five industrial designs; then the technology was tested on real silicon data. After creating the defect Pareto chart from diagnosis reports, the die for FA were selected. FA found defects matching the predicted root causes in six of eight die, a 75% hit rate.

This work shows that cell-aware diagnosis with RCD improves yield ramp and semiconductor quality.

References
“Machine Learning Inside the Cell to Solve Complex FinFET Defect Mechanisms with Volume Scan Diagnosis”. Electronic Device Failure Analysis, Vol. 21, Issue 1, 2019. Copyright © 2019 ASM International. Used with permission. https://static.asminternational.org/EDFA/201902/4/



Leave a Reply


(Note: This name will be displayed publicly)