{"schema_version":1,"research_id":"eoa_inverse_innovation_exp06_external_evaluation_20260803","source_assessment_id":"predictive_residual_processing__nanotechnology:P1:v0","cell_id":"predictive_residual_processing__nanotechnology","search_queries":["site:nist.gov scanning electron microscope semiconductor metrology defects linewidth SEM uncertainty","semiconductor e-beam inspection throughput review layout based defect inspection SEM","layout to SEM image simulation defect detection paper deep learning","KLA e-beam inspection layout database SEM defect review product","site:appliedmaterials.com electron beam inspection SEM review defects throughput product","site:hitachi-hightech.com semiconductor defect review SEM automatic defect classification product","site:semi.org standard defect data wafer inspection SEM review SEMI E142 E134","site:imec-int.com e-beam inspection throughput semiconductor defects SEM"],"sources":[{"source_id":"S1","title":"NIST Study Aims to Improve Utility of the Scanning Electron Microscope","publisher":"National Institute of Standards and Technology","url":"https://www.nist.gov/news-events/news/2025/04/nist-study-aims-improve-utility-scanning-electron-microscope","source_class":"GOVERNMENT_OR_REGULATOR","publication_date":"2025-04-25","accessed_at":"2026-08-03","claims_supported":["SEM is a semiconductor-industry mainstay for imaging very small chip defects and patterns.","Continued miniaturization increases demand for detailed SEM information.","NIST's CHIPS-funded work targets uncertainty in measurements inferred from SEM images."]},{"source_id":"S2","title":"Review SEM — What is a Review SEM?","publisher":"Hitachi High-Tech Corporation","url":"https://www.hitachi-hightech.com/global/en/knowledge/semiconductor/room/manufacturing/review-sem.html","source_class":"OFFICIAL_PRODUCT_DOCUMENTATION","publication_date":"n.d.","accessed_at":"2026-08-03","claims_supported":["Production review-SEM workflows already use defect-coordinate files, image acquisition, storage, automatic review and classification.","Inspection and review systems already detect defects through difference-image comparison with an adjacent die or registered reference pattern.","Inspection may produce several thousand to several tens of thousands of defect candidates, and recipes can select whether all or only some receive review.","Coordinate errors and reference registration are practical workflow limitations."]},{"source_id":"S3","title":"Hitachi High-Technologies Launches New High-Speed Defect Review SEM CR6300","publisher":"Hitachi High-Tech Corporation","url":"https://www.hitachi-hightech.com/file/global/pdf/about/news/2017/nr20170308.pdf","source_class":"COMMERCIAL_FIRST_PARTY","publication_date":"2017-03-08","accessed_at":"2026-08-03","claims_supported":["A commercial review SEM already compares captured defect images with pattern-design data.","The system outputs pattern-shape changes for process optimization and yield enhancement.","Hitachi reported demand for higher defect-capture rates, more images per unit time, failure-area determination and quantification of fine pattern-shape changes."]},{"source_id":"S4","title":"Defect Inspection and Review","publisher":"KLA Corporation","url":"https://www.spts.com/products/chip-manufacturing/defect-inspection-review","source_class":"OFFICIAL_PRODUCT_DOCUMENTATION","publication_date":"n.d.","accessed_at":"2026-08-03","claims_supported":["Commercial optical and e-beam systems already find, identify and classify semiconductor pattern defects.","Commercial inspection uses advanced algorithms and machine learning to suppress or separate nuisance defects.","KLA identifies process engineers and chip manufacturers conducting yield learning, line monitoring and process-excursion resolution as operational users."]},{"source_id":"S5","title":"Defect Detection Approaches Based on Simulated Reference Image","publisher":"arXiv","url":"https://arxiv.org/abs/2303.11971","source_class":"PRIMARY_RESEARCH","publication_date":"2023-03-21","accessed_at":"2026-08-03","claims_supported":["Simulated clean references and difference images have already been evaluated for SEM semiconductor defect detection.","The reported methods span classical difference imaging, supervised learning and unsupervised anomaly detection.","The paper reports better performance from simulated references than real references in its datasets, partly because of reduced noise, geometric variation and registration error."]},{"source_id":"S6","title":"Major Revision Underway for SEMI E142","publisher":"SEMI","url":"https://www.semi.org/en/standards-watch-2026-apr/major-revision-underway-for-semi-e142","source_class":"STANDARD","publication_date":"2026-04-16","accessed_at":"2026-08-03","claims_supported":["SEMI E142 specifies substrate-map data used to report, store and transmit wafer-related map information.","Different process steps can use incompatible XY coordinate systems.","Coordinate alignment is critical for tracking failure points and devices across manufacturing steps, supporting explicit registration and provenance controls."]},{"source_id":"S7","title":"Deep Learning-Based Defect Classification and Detection in SEM Images","publisher":"arXiv","url":"https://arxiv.org/abs/2206.13505","source_class":"PRIMARY_RESEARCH","publication_date":"2022-06-20","accessed_at":"2026-08-03","claims_supported":["SEM noise, shrinking pitches and difficult bridge, collapse and gap classes create false detections and engineering-review burden.","The study used 5,465 raw SEM images but manually labeled only 1,324, illustrating material data-preparation work.","Its ensemble achieved uneven class performance, including substantially lower precision for microbridge and probable nano-gap classes, supporting class-specific safeguards rather than reliance on aggregate accuracy.","The authors identify manual classification and inadequate commercial classification robustness as sources of engineering time."]},{"source_id":"S8","title":"Defect Detection in Photolithographic Patterns Using Deep Learning Models Trained on Synthetic Data","publisher":"arXiv","url":"https://arxiv.org/abs/2505.10192","source_class":"PRIMARY_RESEARCH","publication_date":"2025-05-15","accessed_at":"2026-08-03","claims_supported":["Very small EUV patterning defects can produce false or missed SEM detections.","Scarcity of defect-annotated data, especially for small defects, is identified as a barrier to fab deployment.","Synthetic-data-trained detection transferred imperfectly to real SEM data, reporting 84.6% detection for bridges and 78.3% for breaks, reinforcing the need for real-data validation and protected-class falsifiers."]}],"problem_evidence":{"support":"STRONG","rationale":"Independent government, vendor and research sources show that SEM inspection is important at advanced nodes, candidate volumes and nuisance detections burden review, image interpretation and registration remain uncertain, and small consequential classes remain difficult. The sources establish the general problem, although they do not quantify the candidate's claimed capacity bottleneck for a particular fab, layer or recipe.","source_ids":["S1","S2","S3","S4","S7","S8"]},"stakeholder_evidence":{"support":"MODERATE","rationale":"Commercial first-party sources identify semiconductor manufacturers, process-control engineers and defect-review operations as adopters of design comparison, automated review, classification and nuisance suppression; NIST's CHIPS metrology program is an identifiable funder of improved SEM interpretation. No source expresses demand for the proposal's specific synchronized residual-only representation, raw-audit governance or fallback bundle, and no partner has authorized data access.","source_ids":["S1","S2","S3","S4"]},"prior_art":{"proximity":"SUBSTANTIAL_COLLISION","closest_analogues":[{"name":"Hitachi CR6300 design-data comparison analysis","similarity":"Commercial review SEM compares captured SEM defect imagery with design data and quantifies pattern-shape changes for process control, closely matching the layout-conditioned comparison core.","remaining_difference":"The opened documentation does not establish residual-only transport or review representation, matched predictor checksums, independently sampled raw-tile audits, or automatic full-tile fallback.","source_ids":["S3"]},{"name":"Reference-image difference processing in commercial review SEM","similarity":"Established production practice compares inspected or review images with adjacent-die or registered reference images and uses the difference to find defects while selecting only part of a large candidate set for review.","remaining_difference":"The practice described does not make the signed residual a synchronized reconstructive code or govern suppression with model-version, raw-audit and drift-triggered fallback controls.","source_ids":["S2"]},{"name":"Simulated-reference SEM defect detection","similarity":"Primary research already generates clean reference imagery and uses reference differences in classical, supervised and unsupervised SEM defect detection.","remaining_difference":"The paper evaluates detection performance, not an end-to-end constrained residual channel with independent audit sampling, heartbeats, reconstruction guarantees and engineer-controlled model updates.","source_ids":["S5"]},{"name":"KLA algorithmic nuisance suppression and e-beam review","similarity":"Commercial inspection already uses AI-driven detection, classification and efficient nuisance-defect suppression to allocate inspection and review capacity.","remaining_difference":"The product description does not show reconstruction from prediction plus residual or the proposal's independent suppression-audit and decompression rules.","source_ids":["S4"]},{"name":"Deep-learning SEM defect classification","similarity":"Research systems already detect and localize bridges, line collapses, gaps and microdefects and are designed to reduce manual engineering effort.","remaining_difference":"These are full-image detector/classifier rivals rather than a synchronized reconstructive residual representation.","source_ids":["S7","S8"]}],"distinctive_claim_remaining":"For one fixed layer and SEM recipe, a version-synchronized channel carrying layout-conditioned signed residual patches, coupled to independent random raw-tile audits and triggered full-tile fallback, will reduce total constrained analysis-plus-review resource use versus full-tile compression, fixed layout subtraction and a full-image anomaly detector while remaining noninferior on protected-class recall, engineer disposition decisions and decision-relevant reconstruction. This is falsifiable but not externally demonstrated.","confidence":"HIGH"},"implementation_evidence":{"support":"MODERATE","rationale":"Layout/design comparison, reference differencing, automated review, defect classification, nuisance suppression, coordinate-map exchange and learned SEM detection are independently demonstrated. Technical assembly for an offline shadow replay is credible. Evidence is absent for reliable model synchronization across actual tool services, residual reconstruction tolerances, independent audit sampling rates, fallback behavior, net resource savings and safe performance on rare protected defect classes. Proprietary layout and SEM data are a major implementation dependency.","source_ids":["S2","S3","S4","S5","S6","S7","S8"]},"scores":{"meaningful_impact":{"score":4,"rationale":"If noninferiority is preserved, concentrating review on consequential departures could reduce a documented inspection and engineering burden at advanced nodes. Site-specific impact magnitude remains unmeasured.","source_ids":["S1","S2","S3","S7"]},"stakeholder_pull":{"score":3,"rationale":"Vendors and researchers explicitly address throughput, classification burden, nuisance suppression and process-control use, but no adopter requests this residual-channel architecture or commits data and staff.","source_ids":["S2","S3","S4","S7"]},"incremental_advantage":{"score":2,"rationale":"Design-data comparison, reference differencing, automated triage and nuisance suppression are established. Incremental advantage depends on the untested net benefit of reconstructive residual transport plus governance safeguards.","source_ids":["S2","S3","S4","S5"]},"distinctiveness_plausibility":{"score":2,"rationale":"The complete checksum-audit-fallback bundle was not found in the eight sources, but its technical core substantially collides with existing commercial and research practice. Absence from this bounded search is not world novelty.","source_ids":["S2","S3","S4","S5"]},"technical_implementability":{"score":4,"rationale":"All major computational primitives have close demonstrations, and an offline replay avoids microscope-control changes. Integration, registration and rare-class robustness remain nontrivial.","source_ids":["S2","S3","S5","S6","S7"]},"adoption_authority_feasibility":{"score":2,"rationale":"A process engineer could authorize a read-only internal replay, but no named fab partner, raw-data right, layout-data permission, retention policy or production-interface authority is verified.","source_ids":["S2","S4","S6"]},"evidence_readiness":{"score":2,"rationale":"Published evidence supplies methods and comparators, but decisive evaluation requires proprietary temporally held-out SEM tiles, layout alignment, tool metadata, reviewer labels and measured workflow costs.","source_ids":["S5","S7","S8"]},"safety_net_benefit":{"score":4,"rationale":"Independent raw audits, explicit missingness and full-tile fallback directly address the risk that imperfect predictors suppress rare defects; the need is credible because class performance and transfer to real SEM data are uneven. The safeguards themselves have not been validated.","source_ids":["S6","S7","S8"]},"scalability":{"score":3,"rationale":"Software replay can scale computationally, but every layer, tool, material stack and recipe may require separate registration, calibration, protected-class validation and governance.","source_ids":["S3","S6","S7","S8"]}},"score_confidence":"MODERATE","costs":{"first_evidence":{"band_2026_usd":"50K_TO_250K","scope":"Pre-registration, secure extraction and alignment of one held-out layer/recipe dataset, implementation of three comparators and the residual prototype, blinded review, safeguard simulations and analysis.","confidence":"LOW","assumptions":["Existing raw tiles, layouts, labels and tool metadata are available without purchasing microscope time.","Approximately two to four staff-months across ML, inspection engineering and data engineering.","Compute uses existing workstation or cloud-equivalent resources.","No vendor interface certification is required for offline replay."],"source_ids":["S2","S5","S7","S8"]},"initial_deployment_startup":{"band_2026_usd":"250K_TO_1M","scope":"Secure shadow-pipeline integration for one tool, layer and recipe; provenance and checksum services; audit storage; dashboards; validation; process documentation and operator training.","confidence":"LOW","assumptions":["Existing SEM and yield-management interfaces can export read-only data.","No acquisition-control modification or new SEM hardware is included.","Includes engineering validation and cybersecurity/data-governance work but not fab-wide rollout."],"source_ids":["S2","S3","S6","S7"]},"operational_launch":{"band_2026_usd":"1M_TO_5M","scope":"Qualified limited production shadow operation across multiple wafers on one tool family, including resilient services, formal validation, rare-class test construction, audit/fallback exercises and staffed change control.","confidence":"LOW","assumptions":["Launch remains advisory and cannot autonomously disposition wafers.","Includes vendor/fab integration labor and redundancy but excludes purchase of a new inspection tool.","Multiple recipes require separate calibration and acceptance testing."],"source_ids":["S2","S3","S4","S6"]},"annual_recurring":{"band_2026_usd":"250K_TO_1M","scope":"Model monitoring and revalidation, audit-tile storage and review, incident/fallback testing, software operation, periodic retraining proposals and process-engineer oversight for a limited deployment.","confidence":"LOW","assumptions":["One tool family and a bounded set of layers are maintained.","At least one partial technical FTE plus periodic process-engineer and data-steward effort.","Frequent fallback or recipe proliferation would move recurring cost above this band."],"source_ids":["S2","S4","S6","S7"]}},"verified_pipeline_gates":{"externally_supported_problem":{"status":"YES","reason":"Multiple independent official, commercial and primary sources establish SEM interpretation uncertainty, large review queues, nuisance signals, engineering burden and difficult small-defect classes.","source_ids":["S1","S2","S3","S4","S7","S8"]},"externally_credible_adopter_or_authorizer":{"status":"YES","reason":"Semiconductor process-control and defect-review engineers are an established operational user class, and NIST's CHIPS metrology program is a credible funder of improved SEM interpretation. Interest in this exact architecture is not verified.","source_ids":["S1","S2","S3","S4"]},"distinct_testable_incremental_claim":{"status":"YES","reason":"The remaining claim specifies three comparators, equal resource constraints, noninferiority outcomes and a net-load advantage after audit, synchronization and fallback costs.","source_ids":["S2","S3","S4","S5","S7"]},"bounded_next_evidence_step":{"status":"YES","reason":"A single-layer, single-tool, fixed-recipe offline shadow replay can be preregistered and completed without changing acquisition or wafer disposition.","source_ids":["S2","S3","S5","S7","S8"]},"no_unresolved_safety_or_authority_stop":{"status":"UNCERTAIN","reason":"The offline scope limits physical and disposition risk, but no fab has confirmed authority to use proprietary layouts and raw SEM tiles, data-retention rules, protected-feature definitions or engineer sign-off responsibilities.","source_ids":["S2","S6"]},"credible_cost_scope_and_range":{"status":"UNCERTAIN","reason":"The work packages are bounded, but no opened source supplies labor rates, integration quotations, compute bills or vendor-interface costs. All four bands are resource-equivalent estimates with low confidence.","source_ids":["S2","S3","S6","S7"]}},"next_evidence_step":"With a fab partner, preregister an offline replay of at least 10,000 temporally held-out full SEM tiles from one tool, layer and fixed recipe, including all available labeled defects and deliberate version-mismatch, missing-message and registration-perturbation tests. Freeze models and thresholds. Under an equal compute-and-review budget, compare (A) full-tile conventional compression, (B) fixed layout subtraction, (C) a full-image anomaly detector and (D) the proposed synchronized residual channel. Blind process engineers to method identity and measure protected-class recall, false-positive workload, disposition concordance with unrestricted full-tile review, reconstruction error on a random and risk-stratified raw audit, fallback frequency, latency and total compute, storage, audit, synchronization and reviewer time. Falsify the incremental claim if D misses a protected case surfaced by any comparator, exceeds a preregistered 1-percentage-point noninferiority margin on protected-class recall or engineer decision concordance, leaves any version/missingness ambiguity unresolved, or fails to reduce total constrained-path resource use by at least 25% after all safeguard costs.","blocking_evidence":["No named fab partner has committed temporally held-out raw SEM tiles, corresponding layouts, tool metadata and reviewer outcomes.","The prevalence and decision cost of the claimed full-tile review bottleneck have not been measured for a specific layer and recipe.","Decision-relevant reconstruction tolerance and protected defect classes have not been set by an authorized process engineer.","Rare-class recall, audit sampling power, fallback frequency and failure behavior are unknown.","Net savings after synchronization, storage, audit, fallback, model maintenance and human review are unknown.","Data rights, security controls, retention obligations and production-interface authority are unverified.","No direct cost quotation or measured staff-time baseline supports the four resource bands."],"research_disposition":"PARTNERED_RESEARCH_PROGRAM","world_novelty_boundary":"This evaluation establishes only that the bounded eight-source search found substantial collision in design-data comparison, reference differencing, automated defect review, nuisance suppression and learned SEM detection, while not finding the complete synchronized-residual-plus-independent-audit-plus-fallback bundle. World novelty, patentability, freedom to operate, market size and realized impact remain unmeasured.","arm":"COMPLETE_PROPOSAL_PORTFOLIO","candidate_version":0,"controller_recommendation":{"action":"STOP_EMPIRICAL_RESEARCH_NEEDED","repairable":false,"material_progress_observed":true,"progress_targets":["Secure a named fab partner and written authorization for one bounded offline dataset and reviewer study.","Quantify the baseline full-tile compute and review bottleneck for the selected layer and recipe.","Predefine protected defect classes, reconstruction tolerance, noninferiority margins and minimum net-load reduction with the responsible process engineer.","Run the preregistered four-arm replay with independent random raw-tile auditing and blinded engineering review.","Demonstrate fail-safe behavior for version mismatch, missing messages, registration error, calibration staleness and structured residual drift.","Measure complete resource use and obtain implementation quotations before revising cost gates."],"reason":"Bounded web research can establish the problem, adopters, substantial prior-art collision and technical plausibility, but it cannot determine the remaining incremental claim. That requires proprietary fab data, authorized expert decisions and empirical replay of rare defects and safeguards. Under the controller rule, this evidence dependency requires an empirical-research stop, not further web-only targeted research."},"proposal_index":1}