Semiconductors & electronics
Process data matters when decisions link to yield.
Recipe changes linked to metrology and yield, defect images classified by engineers, and failure analyses can connect an input to an expert decision and a production outcome. That structure is difficult to reproduce from telemetry alone.
01Systems you already run
The raw material is already being recorded.
An assessment starts from the software you run today. It asks what those systems hold, never for the records themselves.
- Manufacturing execution
- Siemens Opcenter Execution Semiconductor, Critical Manufacturing MES, Applied Materials SmartFactory
- Yield management
- PDF Solutions Exensio, KLA Klarity
- Inspection and metrology
- Defect review SEM images, Optical inspection maps, CD and overlay metrology
- Test
- STDF test logs, Wafer sort maps, Final test bins
- Equipment and maintenance
- FDC systems, SECS/GEM event logs, CMMS work orders
- Failure analysis and quality
- FA lab reports, 8D reports, Customer return records
02Dataset opportunities
Where the learnable data sits.
Each opportunity follows a decision from input to outcome. That chain, not the number of records, is what makes data useful for training and evaluating AI.
01
Process changes linked to metrology and yield
Recipe parameters and changes, in-line metrology, electrical test, and final yield joined at the lot or wafer level.
- Recipe change
- In-line metrology
- Electrical test
- Yield outcome
- Engineering disposition
Why it is hard to reproduce
Joining process changes to later yield requires production history and lot or wafer genealogy.
AI use cases
- Process-control and virtual metrology research
- Causal reasoning benchmarks on real interventions
- Agent environments for yield engineering tasks
02
Defect images with engineer classifications
SEM review and optical inspection images labeled by engineers with defect class, likely source, and whether the defect affected yield.
- Inspection flag
- Review image
- Engineer classification
- Root-cause assignment
- Yield impact
Why it is hard to reproduce
An image alone does not establish a defect's cause or yield impact. Engineer labels linked to production results require expert review and the history of the affected wafers.
AI use cases
- Multimodal model training and evaluation
- Few-shot defect classification benchmarks
- Expert-agreement and calibration studies
03
Failure-analysis reports
Internal failures and customer returns traced through electrical characterization and physical analysis to a written root cause and corrective action.
- Failure or return
- Electrical characterization
- Physical analysis
- Root cause
- Corrective action
Why it is hard to reproduce
FA reports can link diagnostic images and measurements to a root-cause judgment and corrective action. Reproducing that chain requires the case evidence and expert analysis.
AI use cases
- Multimodal reasoning evaluation
- Root-cause analysis agents
- Fine-tuning on expert diagnostic narratives
04
Test data and engineering dispositions
Wafer sort and final test results, bin maps, and the engineering decisions on retest, guard-banding, and disposition.
- Test program
- Parametric results
- Bin map
- Engineering decision
- Ship or scrap
Why it is hard to reproduce
Test volume alone does not explain why engineers ordered a retest or changed a limit. Results linked to those decisions and downstream quality can preserve that reasoning.
AI use cases
- Spatial pattern recognition on wafer maps
- Anomaly detection benchmarks
- Decision-support evaluation for test engineering
05
Equipment logs and maintenance decisions
Tool sensor traces and fault events, the technician's diagnosis, the maintenance performed, and whether the tool returned to specification.
- Fault or drift
- Technician diagnosis
- Maintenance action
- Requalification
- Back in production
Why it is hard to reproduce
Sensor traces gain context when linked to a technician's diagnosis, maintenance action, and resulting tool performance. Reproducing that troubleshooting history requires more than raw telemetry.
AI use cases
- Predictive maintenance research
- Troubleshooting agent environments
- Tool-use traces for maintenance workflows
03Volume is not the test
What is usually not valuable on its own.
Lots of data is not the point. A record of an expert deciding, and of what happened next, is.
Raw tool telemetry without events or outcomes
High volume and low signal without the faults, decisions, and results that explain it.
Aggregate yield figures
Yield by product and month says little. The lot and wafer genealogy behind it is what carries signal.
Unlabeled inspection images
Images without classifications or linkage to outcomes provide limited evidence of the engineering decision and its result.
Datasheets, application notes, and standards
Already published and widely indexed.
04Rights and compliance
The sector rules for the dataset.
Rights reviewed before any outreach.
Last reviewed .
How we handle rights and privacy- 01
Trade secrets
Recipes, process windows, and yield learning may be trade secrets. Normalized features, strict use limits, or evaluation in a controlled environment may reduce disclosure risks, but do not eliminate them. Each use needs review for secrecy, contractual restrictions, and export requirements.
- 02
Customer design IP
Customer designs, images, test data, and reports may be subject to IP rights and confidentiality agreements. Holding or generating these records does not establish permission to license them. Customer and third-party rights determine whether specific permission is required for the proposed use.
- 03
Export controls
Some semiconductor technology is controlled under the US Export Administration Regulations (EAR); defense-related technical data may fall under the International Traffic in Arms Regulations (ITAR). Digital access and releases to foreign persons within the US can count as exports. License requirements depend on classification, destination, recipient, end use, and available authorizations or exceptions. Classification and access review must cover the specific package, including derived data.
- 04
Partnership structures
Research or evaluation arrangements with defined use, controlled access, or transformed data may be options. These structures still require rights and export review. Calling an arrangement research, or removing process identifiers, does not itself authorize the disclosure.
05AI use cases
Training examples and evaluation tasks.
Measurements linked to engineering decisions and yield outcomes give industrial AI teams process-control and metrology tasks. Engineer-labeled defect images support classification benchmarks. Failure-analysis narratives connect diagnostic measurements to root causes and corrective actions for reasoning evaluation.
06Questions
What owners in this sector ask first.
- Which records should we assess first?
- Start with engineer-labeled defect images, failure-analysis histories, and process changes linked to metrology and yield. Record the lot or wafer identifiers that connect each decision to its result.
- What can equipment and maintenance records contribute?
- Fault events linked to technician diagnoses, maintenance actions, requalification, and return to production create troubleshooting cases. The sequence explains which action restored performance.
- What makes a failure-analysis report useful?
- Diagnostic images and measurements linked to a root-cause judgment and corrective action form a multimodal reasoning task. Include the case history and the result of the corrective action.
- Is mature-node data still useful?
- For defect classification and maintenance evaluation, engineer labels and linked production outcomes matter alongside the node. Start with the structure of the records and the task to be evaluated.
Find the AI opportunity in your records.
Start with a free assessment of your systems, the decisions recorded, and the outcomes linked to them. You keep ownership.