112965

**N-Detect ATPG and Multiple-Detection Test Pattern Generation** # N-Detect ATPG and Multiple-Detection Test Pattern Generation ## Executive Overview N-detect automatic test pattern generation (ATPG) is a test-generation strategy that requires each fault in the target fault list to be detected by at least N distinct patterns, rather than stopping — as conventional single-detect ATPG does — as soon as any one pattern detects the fault. The motivation is that a fault model such as stuck-at is a simplification: many different physical defects can produce the same stuck-at fault signature, but they do not all sensitize under the same input conditions, and a single detecting pattern reveals only that some defect consistent with that fault model exists, not which physical defect it actually is or whether every electrically plausible variant of it would be caught under different operating conditions. When the generated patterns provide genuinely different sensitization or propagation conditions, multiple detection can increase the odds that patterns collectively cover the real spread of physical defect behaviors hiding behind one fault-model entry, at the cost of a larger pattern set and more test time. This article covers the rationale for multiple-detection test generation, the algorithms used to generate and compact N-detect pattern sets, how the target N value is chosen and tuned in practice, the pattern-volume and test-time economics involved, and how N-detect interacts with related test-quality techniques such as cell-aware test and diagnosis. --- ## Part 1: Why a Single Detection Is Not Always Enough ### The Fault Model as an Equivalence Class A stuck-at, transition-delay, or cell-aware fault model entry is not a single physical defect; it is an equivalence class standing in for every physical defect that would be expected to produce that fault's logical signature. A stuck-at-1 fault on a given net, for example, is consistent with many different underlying causes — a hard short to the supply rail, a resistive short of varying strength, a driver-transistor defect that always pulls the node high regardless of its intended state, or a receiving-gate input defect that always reads the node as high. Single-detect ATPG generates one pattern that sensitizes and propagates *some* member of this equivalence class and then moves on, considering the fault "done." That single pattern was constructed to exercise one specific set of input conditions; it offers no guarantee that a different physical defect in the same equivalence class, which might require different surrounding conditions to manifest or to propagate cleanly to an observable point, would also be caught by that same pattern. ### Detection Path Diversity and Physical Coverage When an N-detect generator deliberately promotes structural diversity, its patterns can sensitize or propagate a fault through different paths, different surrounding signal states, and different routes to different observation points (primary outputs or different scan flip-flops), rather than merely producing N patterns that each happen to detect the fault. Because real defects vary in strength (a resistive bridge partway between a hard short and no defect at all, for instance) and because a defect's detectability can depend on the surrounding circuit's exact state (a marginal defect might only produce an observable logic error under specific noise, coupling, or timing conditions), a fault covered by many electrically diverse patterns is more likely to have at least one of those patterns catch whatever the real underlying defect actually is, compared to a fault covered by only a single, arbitrarily chosen pattern. ### Random and Systematic Defects Beyond the Modeled Fault N-detect's practical value is not necessarily limited to variants of the exact fault being targeted: an unoptimized set of patterns that detect a given modeled fault in genuinely different ways can also incidentally excite and propagate other, unmodeled defect behavior in the surrounding logic. This incidental benefit is not automatic, however, and depends heavily on how the N-detect pattern set is generated and compacted — see the caveat in Part 2 on compaction and untargeted-defect bias, which real published results (Zhang, Thornton, and Dworak's NATW 2014 study) show can work in the opposite direction for aggressively compacted N-detect sets. --- ## Part 2: N-Detect ATPG Algorithms ### Basic N-Detect Generation Loop The conceptual algorithm extends conventional ATPG's per-fault loop: for each fault in the target list, instead of terminating after the first successful pattern, the generator continues producing additional patterns for that same fault — each one required to differ from previous detections in some meaningful way, typically by fixing different "don't care" bits in the test cube or by targeting different propagation paths — until either N detections have been accumulated for that fault or the generator exhausts the practical alternatives available given the surrounding netlist's constraints. ### Structural Diversity Is Not Automatic The defining requirement of N-detect ATPG, as the technique is generally specified, is only that N distinct patterns each detect the fault — it does not, by itself, guarantee that those N patterns differ in any way that matters electrically. A pattern generator satisfying only the literal N-detect requirement could in principle produce N patterns that differ solely in irrelevant, unconstrained bit positions, providing little real benefit over a single detection. Getting genuine sensitization and propagation-path diversity out of an N-detect run requires the generator to be deliberately biased toward it — for example by preferring different values on the fault's don't-care bits or different propagation paths to different observation points across the N patterns — and different commercial and research implementations vary in how strongly, or whether, they do this. This distinction matters in practice: Zhang, Thornton, and Dworak's 2014 NATW study found that N-detect pattern sets optimized (compacted) for pattern-count efficiency can end up *biased against* detecting some untargeted, cell-aware-type defects relative to a less-compacted pattern set — the opposite of the incidental-diversity benefit sometimes assumed for N-detect, and a direct illustration that "N distinct detections" and "N diverse detections" are not the same guarantee. ### Integration with Fault Simulation and Fault Dropping Conventional ATPG uses fault simulation and fault dropping aggressively: once a fault has been detected by any generated pattern, it is dropped from further consideration, which is what keeps single-detect pattern sets compact. N-detect ATPG modifies this bookkeeping to track, per fault, how many distinct detections have been accumulated so far, dropping a fault only once its target count is reached; this requires the fault simulator to continue evaluating already-partially-detected faults against every subsequently generated pattern (since a pattern generated primarily to target a different fault may incidentally also detect an under-detected fault), which increases fault-simulation cost relative to a single-detect flow. ### Static and Dynamic Compaction Trade-offs Static and dynamic pattern compaction — merging compatible fault requirements into shared patterns — apply to N-detect generation just as they do to single-detect ATPG, but with a modified objective: rather than compacting to the fewest patterns that detect every fault at least once, N-detect compaction must preserve enough diversity across the pattern set that each fault still accumulates N genuinely distinct detections. Overly aggressive compaction that merges too many fault requirements into too few patterns can undermine the diversity N-detect is meant to provide, so N-detect flows generally accept a less compact pattern set than single-detect ATPG would produce for the same fault list, by design rather than as an unavoidable side effect. --- ## Part 3: Choosing the Target N Value ### N as a Tunable Test-Quality Knob There is no single universally correct value of N; it is a test-program design parameter traded off against pattern volume and test time, similar to a target fault-coverage percentage. Published experimental evaluations have generally studied specific, modest values rather than establishing broad tiers — for example, Yu Huang's ISQED 2006 pattern-set optimization study ("On N-Detect Pattern Set Optimization") evaluates N=3 and N=5 as its experimental targets. The right value of N for a given design depends on its defect-density expectations, its reliability requirements, and its test-time budget, and should be set from that design's own trade-off data (or from a specific published study with a comparable design and process) rather than from a generic recommended range, since no broadly applicable tiering of "typical" N values is established in the literature. ### Diminishing Returns and Saturation It is plausible on structural grounds that the marginal test-quality benefit of each additional required detection per fault diminishes as N grows — the number of genuinely distinct, structurally meaningful sensitization-and-propagation paths available for a given fault in a given netlist is finite, so at some point additional required detections can only be satisfied by decreasingly diverse patterns. This reasoning is not, however, backed here by a study that directly measures a diminishing-returns curve across a range of N values; where a saturation point falls for a given design, and whether it is pronounced enough to matter for pattern-budget planning, is best established from that design's own fault-coverage and pattern-count trade-off data rather than assumed from the general argument above. ### Selective N-Detect Application Rather than applying a uniform N across an entire fault list, some flows apply a higher N selectively — to faults on nets or in cells identified as higher-risk from yield-learning or failure-analysis data, or to faults associated with critical timing paths — while accepting single-detect or low-N coverage elsewhere in the design. This targeted approach, conceptually similar to the selective application of cell-aware test discussed elsewhere in this knowledge base, aims to concentrate the added pattern-volume cost where it is expected to deliver the most defect-screening benefit rather than spreading it uniformly across a fault list of uneven risk. --- ## Part 4: Pattern Volume, Test Time, and Cost Trade-offs ### The Core Economic Trade-off Every additional required detection per fault adds test patterns (though sub-linearly, due to compaction and the incidental-detection effect described in Part 1), and every additional pattern adds automated test equipment (ATE) time per die at production volume. Test program design has to weigh the marginal defect-screening improvement N-detect provides against this direct, easily quantified per-die cost increase, which is why N-detect adoption decisions are typically driven by the same reliability-requirement and quality-target reasoning that governs adoption of other coverage-enhancing techniques rather than applied unconditionally. ### Relationship to Outlier and Statistical Screening N-detect pattern sets are sometimes discussed alongside statistical outlier-screening techniques (such as part-average testing or current-signature-based outlier detection) as complementary approaches to catching defects that a purely pass/fail functional or structural test might miss: N-detect increases the odds that a structural pattern set directly excites and catches a marginal defect, while statistical outlier methods catch parts whose electrical behavior deviates from the population without necessarily identifying which specific fault or defect is responsible. The two approaches address the same broad reliability-screening goal — reducing field failures from defects that pass conventional test — through different mechanisms, and are not substitutes for each other. ### A Potential Diagnosis Benefit Because N-detect pattern sets exercise a fault through multiple distinct sensitization and propagation conditions — to the extent the pattern set was actually generated with genuine diversity, per the caveat in Part 2 — a failing part's response pattern across such a set could plausibly carry more information for diagnosis than a single-detect set would: which of the N patterns for a given fault failed, and which passed, could in principle narrow the plausible defect location and behavior more precisely than a single pass/fail result per fault. This is a reasonable mechanism-level expectation rather than a result established by a diagnosis-specific study cited here, and it should be treated as a potential benefit to be verified against a specific design's diagnosis data rather than an assumed outcome of adopting N-detect. --- ## Part 5: Interaction with Other Test-Quality Techniques ### N-Detect and Cell-Aware Test N-detect and cell-aware test address related but distinct gaps in pin-level fault modeling: cell-aware test adds an entirely new fault model targeting defect locations invisible to pin-level fault models at all, while N-detect increases the number of independent sensitization conditions used to detect faults within whatever fault model (pin-level or cell-aware) is already being targeted. The two techniques can be applied together — generating multiple, structurally diverse detections for each cell-aware fault, not just for pin-level faults — for designs whose reliability requirements justify the combined pattern-volume cost of both techniques. ### N-Detect and Timing-Aware Delay Test For transition-delay and small-delay fault models, N-detect can be combined with path-diversity requirements — specifically requiring that the multiple detections for a given delay fault be routed through electrically distinct timing paths with different amounts of slack — which increases the odds of catching a marginal delay defect whose detectability depends on which specific path timing margin it erodes, an extension of the same equivalence-class reasoning applied to delay rather than logic faults. --- ## Part 6: Fault Coverage Reporting and Signoff ### N-Detect Coverage Metrics N-detect coverage can be reported not as a single blended percentage but as a distribution: what fraction of the target fault list achieved the full target N detections, what fraction achieved a partial count, and what fraction remains at a single detection or undetected, since a design with 95% of faults detected the full target N times represents meaningfully higher test quality than one with 95% of faults detected merely once, even though both might report a superficially similar single-detect fault-coverage percentage. Whether a given test program actually adopts a distribution-based signoff criterion, and what floor it sets, is a program-specific decision rather than an established industry norm cited here. ### Practical ATPG-Untestable and Detection-Limited Faults Just as some faults are entirely ATPG-untestable under single-detect generation (due to unreachable logic states or redundant circuit structure), some faults that are single-detect testable are nonetheless detection-limited under N-detect targets — the netlist genuinely does not offer N structurally distinct sensitization-and-propagation combinations for that fault, regardless of how much ATPG effort is applied. Distinguishing genuinely detection-limited faults from faults that simply require more sophisticated pattern generation is an ongoing practical concern in N-detect tool tuning, analogous to the corresponding distinction in cell-aware ATPG. --- ## Part 7: Production Deployment Considerations ### Fault List Scope N-detect targets are commonly applied across the full stuck-at and transition-delay fault lists for a design, though as noted in Part 3 some programs apply higher N selectively to a risk-prioritized subset rather than uniformly. Deciding scope requires the same kind of risk-prioritization input (yield-learning data, criticality of specific circuit blocks, reliability requirements for the end application) needed to make any selective-coverage decision economically sound rather than arbitrary. ### Compatibility with Standard Scan-Based Test Infrastructure Like cell-aware test, N-detect ATPG generates patterns applied through the same scan-chain shift/capture infrastructure already present for conventional structural test, requiring no additional design-for-test hardware. This makes N-detect adoption primarily an ATPG tool configuration and test-time budgeting decision rather than a design-for-test architecture change, which lowers the practical barrier to adopting it relative to techniques that require new on-chip test structures. ### Tool Support and Configuration N-detect generation is a standard configuration option in mainstream commercial ATPG tools, typically exposed as a target-N parameter alongside conventional fault-coverage and pattern-count targets, and is commonly combined with other ATPG configuration choices (compression ratios, at-speed launch-capture settings, and fault-model selection) as part of an overall test-pattern generation strategy tuned per design rather than applied as a single fixed tool default across all products. --- ## Summary: N-Detect as a Physical-Diversity Complement to Fault-Model Coverage N-detect ATPG addresses a limitation inherent to any fault-model-based test strategy: a fault model entry is an equivalence class standing in for a range of physical defect behaviors, and a single detecting pattern demonstrates only that some member of that class is caught, not that every member would be. Requiring multiple detections for each targeted fault can increase the odds of catching whichever specific physical defect variant is actually present and can, in principle, provide a richer per-fault pass/fail signature useful for diagnosis — but only to the extent the pattern set was generated with genuine sensitization and propagation-path diversity, which is not automatic and which published research (Zhang, Thornton, and Dworak, NATW 2014) shows can actually run in reverse for pattern-count-optimized sets, biasing them against some untargeted defects. Its cost — additional pattern volume and test time — makes it, like cell-aware test, a tunable test-quality lever rather than a fixed requirement: the appropriate target N, whether to apply it uniformly or selectively, and how much to prioritize genuine pattern diversity over pattern-count efficiency, are design- and reliability-requirement-specific decisions best grounded in a design's own trade-off data and in the specific published studies most comparable to that design, rather than in a general industry rule. --- ## Selected Published N-Detect Studies The table below summarizes specific, citable experimental results rather than a generalized recommendation — each row is one study's own design, N value, and reported outcome, not a broadly applicable tier. | Study | Study Configuration | Design/Context | Reported Result | |---|---|---|---| | Venkataraman, Sivaraj, Amyeen, Lee, Ojha & Guo, VTEST 2004 | Experimental N-detect evaluation | Microprocessor | Experimental study of N-detect scan ATPG patterns on a real processor netlist | | Amyeen, Venkataraman, Ojha & Lee, ITC 2004 | Production N-detect evaluation | Microprocessor (production design) | Quality evaluation of N-detect scan ATPG patterns on a processor | | Huang, ISQED 2006 | N = 3, N = 5 | Pattern-set optimization | Pattern-count-efficient generation of N-detect sets evaluated at these specific N values | | Geuzebroek, Marinissen, Majhi, Glowatz & Hapke, ITC 2007 | Embedded multi-detect configuration | Production defect-detection evaluation | 2.3%-4.7% additional defective-device detection reported for the specific embedded multi-detect approach evaluated — a single study's result, not a general figure | | Zhang, Thornton & Dworak, NATW 2014 | Compacted vs. uncompacted N-detect sets | Untargeted (cell-aware-type) defect coverage | Found that pattern-count-optimized N-detect sets can be biased *against* detecting some untargeted defects relative to less-compacted sets | --- **Sources**: S. Venkataraman, S. Sivaraj, M.E. Amyeen, Sangbong Lee, A. Ojha, and Ruifeng Guo, "An Experimental Study of N-Detect Scan ATPG Patterns on a Processor," IEEE VLSI Test Symposium (VTS), 2004, DOI: 10.1109/VTEST.2004.1299221; M.E. Amyeen, S. Venkataraman, A. Ojha, and Sangbong Lee, "Evaluation of the Quality of N-Detect Scan ATPG Patterns on a Processor," IEEE International Test Conference (ITC), 2004, DOI: 10.1109/TEST.2004.1387328; Yu Huang, "On N-Detect Pattern Set Optimization," IEEE International Symposium on Quality Electronic Design (ISQED), 2006, DOI: 10.1109/ISQED.2006.94; Jeroen Geuzebroek, Erik Jan Marinissen, Ananta Majhi, Andreas Glowatz, and Friedrich Hapke, "Embedded Multi-Detect ATPG and Its Effect on the Detection of Unmodeled Defects," IEEE International Test Conference (ITC), 2007, DOI: 10.1109/TEST.2007.4437649; Fanchen Zhang, Micah Thornton, and Jennifer Dworak, "When Optimized N-Detect Test Sets Are Biased: An Investigation of Cell-Aware-Type Faults and N-Detect Stuck-At ATPG," IEEE North Atlantic Test Workshop (NATW), 2014, DOI: 10.1109/NATW.2014.15. Claims not directly traceable to one of these five studies are presented as general design-for-test reasoning rather than as an established empirical result, and are marked accordingly in the text above.