Sponsored Content material: A brand new method to resolve systematic circuit failures and enhance yield

Bettering yield is vital to decreasing manufacturing price and maximizing earnings in a extremely aggressive semiconductor market. Whereas giant yield limiters are addressed early within the course of throughout yield ramp, delicate systematic structure patterns proceed to trigger bodily defects all through quantity manufacturing.

The R&D groups from Siemens’ Tessent group and PDF Options developed a brand new method that mixes sample evaluation—FIRE from PDF Options—with quantity logic scan prognosis—root trigger deconvolution (RCD) from Siemens EDA. The FIRE sample evaluation engine determines suspect teams of comparable structure patterns (sample households), the Tessent RCD diagnosis-based machine studying software program then produces defect Paretos. These yield Paretos embrace particular structure sample households as distinct root causes and present an general improve in defect Pareto accuracy from ~70% to ~90%. Siemens EDA and PDF authors described it in an invited paper on the 2023 IEEE SPIE Superior Lithography + Sample convention and summarize the method right here.

Fixing the issue of systematic failures

Failures throughout manufacturing might be categorised as random or systematic, the place systematic failures imply that sure traits, just like the native geometrical neighborhood, decide the prevalence of failures. There are a couple of methods to focus on these neighborhood influences, however new or surprising systematic failures can nonetheless be the primary contributor to yield loss even at mature phases of contemporary applied sciences.

The following step ahead in yield enchancment is to determine these systematic failures. We have to determine failing patterns and to quantify their influence, however the strategies obtainable—SEM imaging and bodily failure evaluation—are time consuming and dear, want an a-priori info of the areas and may’t be utilized to a big sufficient quantity to be able to quantify the influence on product yield. We want a way that may quantify the influence of a sample and may present areas.

Right here’s how we deal with this downside:

• Analyze the geometry of a product to outline all areas the place a failure can occur and construct the patterns round these areas.
• Group the patterns into households, as proven in determine 1.

Determine 1: Sample households on a flat structure after extraction and sample evaluation.

Utilizing highly effective instruments: logic scan prognosis software program

Scan patterns generated by an automated check sample technology (ATPG) software are the defacto customary in manufacturing check of digital designs. These scan patterns comprise stimulus and known-good responses calculated by simulating the logical design. The manufactured design is examined by making use of the design’s ATPG-generated scan patterns utilizing automated check gear (ATE), which captures the failing responses. This fail knowledge—scan patterns together with the logical and bodily view of the design—are utilized by scan prognosis software program to supply a report, a course of illustrated in determine 2.

Determine 2: Typical scan prognosis circulate.

The everyday layout-aware, cell-aware scan prognosis gives info on location and bodily properties that designate the failures. The reviews information failure evaluation and yield evaluation. Quantity prognosis, which is a big set of scan prognosis, is especially efficient the check failure knowledge is transformed right into a set of bodily suspect areas for every failing die. For yield evaluation, machine studying strategies like root-case deconvolution (RCD) assist clarify a inhabitants of quantity scan prognosis reviews.

The way to convey sample household info into RCD evaluation

We prolonged RCD to study structure feature-driven systematic yield loss points by combining it with sample household knowledge obtained from analyzing the design structure. The sample household info contains of a listing of factors with a location (x, y and layer) within the bodily structure that may be a potential defect location. Every level is given a sample household and whether or not it might be an open or quick.

The general circulate to import this info into quantity prognosis setup is proven in determine 3. Step one matches the purpose areas to logical nets within the gate-level netlist. This can be a one-time setup step that offers the Tessent prognosis structure database a map of each web phase and bridge location within the design to a listing of potential defect location factors on it.

Determine 3: Format sample, prognosis and RCD circulate.

This mapping info is then used throughout prognosis to annotate the prognosis suspects with structure sample household info. Lastly, when studying from quantity prognosis outcomes utilizing RCD, the structure sample households get added to the listing of potential defect root causes that RCD should decide from when estimating the defect Pareto for a given failing die inhabitants.

Outcomes of mixing structure sample households and RCD

There might be tons of of 1000's of structure sample households in a contemporary design. Studying a defect Pareto from so many potentialities considerably will increase the possibilities of converging on a incorrect resolution. To beat this, RCD will intelligently filter out extremely unlikely options and trim ‘tail finish’ structure sample households which can be unlikely to be main causes of yield loss.

We examined this technique in a simulated atmosphere that creates populations of failing die primarily based on a given defect Pareto. For each defect Pareto a couple of hundred faulty dies had been created by simulating injected defects after which diagnosing the ensuing check failure logs. Lastly, the quantity prognosis outcomes had been fed into the improved RCD studying and the estimated Pareto was in contrast towards the beginning defect Pareto to measure accuracy in figuring out injected structure sample households.

We repeated this experiment for seventy-two totally different populations to measure common accuracy throughout totally different defect situations. The outcomes are documented in determine 4. This chart reveals the common proportion of injected root causes appropriately discovered is 95%.

Determine 4: Proportion of injected root causes appropriately discovered.

Determine 5 reveals that particularly for design systematic structure sample household root causes, on common 3.75 of the 4.17 design systematic root causes had been appropriately discovered. This knowledge demonstrates the advantage of our methodology in figuring out beforehand unknown design systematics from manufacturing check knowledge.

Determine 5: Accuracy of figuring out injected design sample systematic root causes for numerous populations with totally different defect combos.

Abstract

Ramping and sustaining excessive yielding trendy superior course of applied sciences continues to problem the business. Our new methodology helps uncover delicate defect modes the place sample systematics work together with designs to supply yield losses. The mixture of efficient structure sample evaluation (FIRE), logic scan prognosis and RCD reveals promise to uncover such points with out the necessity for in depth failure evaluation. Utilizing structure patterns and logic check failure knowledge, the simulation outcomes detailed above present excessive confidence (95% settlement) that this technique may also help in actual manufacturing designs. We are going to proceed working with companions and prospects to validate the methodology with excessive quantity knowledge units from manufacturing manufacturing layouts and prognosis. Study extra about quantity prognosis on the Siemens web site.

Authors

Manish Sharma and Jayant D’Souza, Siemens EDA

Hans Eisenmann and Thomas Zanon, PDF Options

Jayant D’Souza, Siemens EDA