August_EDFA_Digital

edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 20 NO. 3 54 GUEST COLUMNIST ARTIFICIAL INTELLIGENCE IN ELECTRONIC DESIGN AUTOMATION ASSISTING PHYSICAL FAILURE ANALYSIS Yu Huang, Mentor, A Siemens Business yu_huang@mentor.com P hysical failure analysis (PFA) has relied exclusively on advanced FA equipment in the past. However, without the assistance of Electronic Design Automation (EDA) so¨ware, it has become very diƒicult to narrow down the location of defects on a big chip. State- of-the-art diagnostic EDA tools can report suspects within a very small region on a layout and can even pinpoint to a few transistors. Without such diagnostic EDA tools, PFA engineers may have a hard time initiating failure analy- sis. Typically, PFA engineers are not concerned with the techniques behind EDA tools, as long as the equipment can report the correct location and suspect for them to dive in. However, understanding the recent enhancements of EDA tools from a high level may help PFA engineers achieve their targets more effectively and efficiently. Among the recent enhancements in EDA diagnostic and yield learning tools are artificial intelligence (AI) technolo- gies such asmachine learning. This column reviews three AI technologies applied in state-of-the-art EDA diagnostic and yield learning tools and suggests how PFA engineers can benefit from them. AI-ASSISTED ELECTRICAL DIAGNOSIS In the past, simulation-based electrical diagnosis was widely used in EDA diagnostic tools. Physical defects are modeled as logical stuck-at or transition faults. Candidate faults were injected into the netlist and fault simulation was run. If the simulated failures “match” the silicon fail- ures observedon the die tester, the injected logical fault(s) are reported as suspect(s). PFA engineers are then guided by the diagnosis report to dive in to the area containing the suspect(s). The criteria for a “match” varies from tool to tool. Essentially, it measures the similarity between logical failure models and observed electrical defective behavior. It is relatively accurate as long as the logical faults can model the real silicon defects. However, this scenario changed when new technology nodes moved to nanometers, FinFET and other advanced technologies became popular, and more complicated cell libraries emerged. Today it is very diƒicult and expensive tomodel silicon defects accurately, which leads to a reduced PFA success rate, if still using simulation-based EDA tools. Therefore, the current trend is to use AI to bridge the gap between fault model and real defect behavior, to tolerate inaccuracy and noises. One example is a paper recently published in the Proceedings of Asian Test Symposium. [1] Unsupervised machine learning techniques are used to partition scan cells on one faulty scan chain intomultiple groups, where the defective cells are locatedat the bound- ary between two adjacent groups. The diagnostic score is calculated based on Bayesian inference equations. PFA engineers should interpret these scores as probabilities of the reported scan cell(s) being defective. AI-ASSISTED SYSTEMATIC DEFECT ROOT CAUSE IDENTIFICATION Root cause deconvolution (RCD) [2] is a novel yield analysis technologybasedonmachine learningmethodol- ogy. It works by analyzing many layout-aware diagnosis reports together to identify the underlying root cause distributionmost likely to explain a particular set of diag- nosis results. It has two benefits for PFA engineers: First, it can identify the failing die that most likely represent the systematic defect of interest. PFA on these failing die has amuch higher success rate identifying systematic defects than PFA on randomly selected failing die. In that sense, RCDworks as a “virtual PFA,” with the ability to determine defect distribution for a population of failing devices before real PFA is performed. Second, the RCD results can be back annotated to the individual diagnosis suspect. Where layout-aware diagnosis points to a segment of interconnect, RCD can isolate a particular root cause in that segment, such that diagnostic resolution is enhanced. Therefore, it will increase the PFA relevance and success rate dramatically, while reducing PFA cycle time.

RkJQdWJsaXNoZXIy MjA4MTAy