Synthetic Data for Manufacturing AI: RFP and 90-Day PoC Guide

The hard part of a synthetic data manufacturing program is not generating a large number of images. It is contracting what shortage in reality will be filled by which synthetic conditions, and deciding which claims must still be proven with real factory data. In AI visual inspection and robot or object perception, rare defects, unsafe poses, occlusion, glare, and post-changeover states are difficult to collect. Yet a simulation-heavy program can hide the domain gap between generated and operating conditions. This guide helps manufacturing buyers specify the data contract, real–synthetic mix, RFP, 90-day PoC, stop rules, and acceptance evidence.

Executive answer: buy an evidence-producing data process

Synthetic data is not automatically a replacement for real data. Treat it as a controlled supplement for conditions that are hard, slow, unsafe, or expensive to observe. A procurement-ready program needs six commitments:

1. Freeze the task, cost of errors, operating envelope, and exclusions.
2. Separate real data for training, tuning, and final acceptance; keep the final set away from the generator team.
3. Contract the factors, distributions, labels, provenance, and generation version.
4. Compare real-only, synthetic-only, and mixed training against the same real holdout.
5. Set floors by defect, asset, lighting, lot, and operating slice instead of accepting one average score.
6. Predefine when to stop if the domain gap, operating burden, or total cost does not improve.

The NIST 2026 roadmap for AI and machine learning in smart manufacturing highlights industrial-data complexity, data management, heterogeneous sensing and control integration, and the need for trustworthy, explainable, reliable operation. The roadmap is not a certification of any synthetic-data product. It does support a system-level procurement principle: evaluate the data, integration, measurement, and operation—not just a model score.

Classify the shortage before generating defect images with AI

“We do not have enough defect images” is too broad for an RFP. Split the shortage into distinct procurement problems.

Synthetic data mainly addresses the need for examples. It does not by itself fix camera vibration, dirty lenses, trigger mismatch, lighting decay, operator placement, process changes, or inconsistent labeling. Maintain separate issue lists for data coverage and for equipment or operational readiness.

Choose the generation method by control need

There are four common families.

• 3D or physics-based rendering: models CAD, material, light, camera, sensor, and defect geometry. It provides controllable pose, occlusion, and geometric labels, but requires asset creation and reality matching.
• Image composition: overlays a defect or foreign object onto real backgrounds. It can be fast, but unrealistic edges, shadows, reflections, and defect locations can teach shortcuts.
• Generative models: create candidates from text, masks, or reference images. They can increase appearance diversity, but provenance, geometry, reproducibility, rights, and frequency still need control.
• Pseudo defects: modify normal images to produce anomaly cues. They can help anomaly detection or few-shot adaptation when true defects are absent, but need not represent the real failure mechanism.

BladeSynth, published in Scientific Data in 2025, is a research dataset for aero-engine blade inspection. It uses physically based rendering and domain randomization to produce 12,500 high-resolution images with segmentation masks across four defect classes: corrosion, notch, dent, and scratch. This validates a specific blade asset, defect set, and rendering workflow; it does not establish performance for molded plastic, welding, or food packaging. PDMCNet, published in Scientific Reports in 2026, synthesizes pseudo defects from normal support images to calibrate few-shot segmentation of unseen categories. It reports mIoU of 36.03% in a one-shot setting and 38.98% in a five-shot setting on Industrial-5i. Those figures belong to that benchmark and setup, not to an arbitrary production line.

Architecture for simulation data in manufacturing AI

NVIDIA’s Omniverse Replicator documentation separates the simulation-to-reality domain gap into appearance gap and content gap. Appearance gap covers pixel-level differences such as materials, surface detail, lighting, rendering, and sensor behavior. Content gap covers differences in object type, count, placement, and context. A factory RFP should expand those two categories into a domain-gap register.

Do not specify the generator as “a system that produces realistic pictures.” Specify it as a data-production process that outputs controlled factors, evidence for their ranges, source assets, generation code or model version, random seeds, labels, and QA results. An image that cannot be traced or regenerated is hard to audit when a defect is added or the model is updated.

The synthetic data manufacturing contract

Require this contract before a volume target.

The distribution rationale is critical. “Randomly generate scratches from 0 to 20 mm” is incomplete. Why is that range plausible? Does the process or quality standard support it? Should sampling be uniform? How many boundary and extreme cases are required? Oversampling rare defects may improve learning while changing probability calibration and threshold behavior. Record the training sampling distribution separately from the prevalence used for acceptance and operations.

Compare real-only, synthetic-only, and mixed data

There is no universal best percentage of synthetic data. Do not lock a ratio before the baseline. Compare at least three tracks under the same model family, compute budget, split, and evaluation code:

1. Real-only: all currently available real data.
2. Synthetic-only: tests the sim-to-real capability and reveals failure boundaries.
3. Mixed: real plus synthetic, with more than one ratio.

Also distinguish pretraining on synthetic followed by fine-tuning on real from mixing both sources in each batch. The 2025 study *Fully-Synthetic Training for Visual Quality Inspection in Automotive Production* used domain randomization and reported that, in three real inspection scenarios, an object detector trained only on synthetic images could outperform models trained on real images. It is evidence for the authors’ three scenarios, model, and pipeline—not a claim that every plant can eliminate real data. Its practical value is to justify a synthetic-only comparison arm and a careful real holdout.

Prevent leakage in the data split

Random image-level splitting can place the same part, adjacent video frames, or near-duplicate renderings from one seed family in training and testing. Split by an appropriate higher-level unit: lot, date, machine, mold, part serial, acquisition session, CAD family, or generation-seed family.

The final acceptance set should contain:

• real images frozen before the PoC and hidden from generation and training teams;
• a new lot, material state, lighting state, machine, or camera instance;
• difficult normal samples such as glare, contamination, texture, and fixtures;
• critical defects and confusing non-defects;
• a later time period to detect degradation under future conditions.

Synthetic samples can be used to test generator quality, but they do not prove performance on the plant. The primary Go/No-Go decision should use isolated real data and a line trial.

RFP requirements for synthetic-data visual inspection

An RFP should request the artifacts needed to diagnose a miss, not merely a number of delivered files.

1. Scope and exclusions

Identify the line, product families, defects, cameras, takt time, decision point, and downstream action. List excluded defects and equipment. A broad sales demonstration must not become an implied production guarantee.

2. Assets and rights

State the rights for CAD, textures, real images, generated assets, model weights, prompts, adapters, and third-party materials. Cover derived works, cross-border transfer, vendor model training, retention, and deletion after the contract.

3. Reproducible generation

Require software and model versions, code, configuration, seeds, dependencies, and hardware in a manifest. FAT should demonstrate regeneration of sampled items. For nondeterministic generators, accept a documented distribution and quality gate rather than demanding pixel-identical output.

4. Real-data measurement

Record illuminance, camera distance, focus, exposure, lens, background, conveyor speed, part pose, and machine state. A factor that has not been measured cannot have a defensible randomized distribution. Where images cannot leave the factory, compare an on-premises method that exports approved statistics rather than raw images.

5. Dataset quality assurance

Automate checks for duplicates, invalid labels, impossible defect locations, empty masks, class mismatch, artifacts, watermarks, text, and confidential objects. Quality and manufacturing engineers should then review a stratified sample for physical plausibility and inspection criteria—not merely visual attractiveness.

6. Performance floors

Report task-appropriate measures such as precision, recall, F1, mAP, mIoU, or pose error by defect, product, asset, illumination, pose, and size. At line level, measure escaped defects, false rejects, reinspection, manual-review rate, inference time, and downtime.

7. Update and exit conditions

Define when the gap register is reviewed, how new real defects are incorporated, and which changes trigger retesting. Contract delivery of data, assets, code or configuration, manifest, evaluation output, known limitations, and deletion evidence at exit.

For the broader inspection system and camera-to-operation design, see our AI visual inspection implementation guide. This article deliberately concentrates on procurement and acceptance of the data layer.

A 90-day synthetic data PoC

Ninety days is an example, not a promise. CAD preparation, real-defect collection, equipment changes, safety work, or customer approval can require more time. Manage gates by evidence rather than calendar alone.

Days 0–15: value hypothesis and data contract

• document the process, current inspection, loss of errors, and annual volume;
• have Quality approve the defect ontology and acceptance boundary;
• inventory real data by lot, asset, and time;
• choose gap hypotheses, generation methods, rights, and security controls;
• freeze the isolated real acceptance set and evaluation code;
• run the real-only baseline.

Do not start generation if inspectors cannot agree on the defect, if real camera conditions cannot be recorded, or if an independent holdout cannot be protected.

Days 16–35: minimum generator and data QA

Build a minimum pipeline for one product, one camera, and one or two defect types. Produce the manifest and labels, run automated QA, and have Quality review a stratified sample. Scale generation only after shortcut cues and physically impossible samples are addressed.

Days 36–55: real-only, synthetic-only, mixed comparison

Use the same model, training budget, and inference conditions. Review slice-level performance and error types, not only the aggregate. Synthetic data may improve common conditions while increasing false alarms on reflective surfaces. Do not let an average hide that regression; return it to the gap register.

Days 56–70: line shadow test

Run alongside the current inspection without controlling disposition. Observe correlated video frames, takt, network delay, camera reconnection, product changeover, cleaning, and lighting decay. Record human review time and the operational response to false rejects.

Days 71–85: closed-loop correction

Classify errors as real-data shortage, synthetic-distribution shortage, label disagreement, equipment condition, or model limit. Change one causal group at a time. Record whether improvement came from more synthetic data, more real images, optical changes, or threshold adjustment.

Days 86–90: accept or stop

Run the frozen real set and selected production shifts. Decide Go, conditional Go, extension, or Stop. Accept the data manifest, rights, monitoring, update plan, rollback procedure, training, and known limitations—not just model weights. Combine this plan with our AI PoC exit-criteria guide to prevent indefinite pilot extension.

Acceptance: process loss and slice floors before averages

The right metric depends on classification, detection, segmentation, or pose estimation. Put four levels into the RFP.

Do not select a universal threshold such as “98% recall” without the current inspection, customer requirement, redundancy, sample size, and confidence interval. Detecting all 20 critical examples does not tightly estimate future recall. Report the observed count and uncertainty, extend the shadow period where needed, and retain process safeguards until evidence is sufficient.

The acceptance evidence pack

Bind the following into one controlled release:

• requirement ID, process, risk, and accountable owner;
• real and synthetic manifests, provenance, licence, and hashes;
• factor table, random seeds, generator code or model version;
• split evidence and duplicate test;
• real-only, synthetic-only, and mixed comparison;
• slice results, confusion matrix, and representative errors;
• line shadow logs, disposition, takt, and downtime;
• domain-gap register and unresolved items;
• change history, retest, and approvers;
• model and data rollback procedure with a restoration test.

A PDF report alone is insufficient. Require a machine-readable manifest that traces each result to data, model, configuration, and execution logs. This prevents silent dataset replacement while the same report remains in circulation.

Stop rules for a disciplined manufacturing AI PoC

Stop criteria protect investment. Agree examples such as these before the PoC:

• the critical-defect floor on the isolated real set is missed in two consecutive gated evaluations;
• doubling synthetic volume does not deliver the pre-agreed minimum improvement on real data;
• a critical machine, product, or illumination slice remains weak without a credible correction;
• provenance, rights, deletion, or regeneration cannot be audited;
• inspector agreement is below the threshold, so the target truth is not stable;
• takt, manual review, or false rejects exceed the process limit;
• recurring collection, regeneration, training, and acceptance cost exceeds the cap;
• vendor lock-in prevents delivery of assets, manifests, or evaluation evidence.

Allow an extension only where the cause is identified, a credible change can address it, and a time, cost, and reacceptance condition are written. “The model may improve with more training” is not a sufficient extension case.

Transparent ROI and TCO assumptions

Synthetic data can reduce some collection and labeling, enable pre-production work, and reproduce unsafe conditions. It also adds CAD and material work, generator QA, domain-gap analysis, real acceptance, and continuing updates.

The following is a hypothetical illustration, not a market price or performance guarantee.

The simplified baseline data cost is 20,000 × 300 = THB 6,000,000. The synthetic option is 8,000 × 300 + 2,000,000 + 900,000 = THB 5,300,000, giving a hypothetical first-year difference of THB 700,000. If 12,000 real examples are still required, the option becomes THB 6,500,000 and loses the cost advantage. Therefore the real-data reduction rate is a PoC variable, not guaranteed savings. Add quality-escape or launch-time benefits only after measuring them.

TCO should include compute, render farm, cloud transfer, asset creation, generator review, MLOps, licences, storage, factory capture, reacceptance, training, and audit. An image-unit price can hide most recurring work.

Governance for Thailand manufacturing sites

ETDA’s 2026 “Driving Trust AI Governance” direction emphasizes guidelines and toolkits, impact assessment, testing, and AI that is safe, transparent, fair, and responsible. It does not certify a synthetic-data method. It does provide local policy context for keeping provenance, owners, risks, tests, and evidence in a Thai factory AI project.

Practical procurement questions include:

• Can CAD, real images, and generated assets move between the Thai site, headquarters, and vendor?
• What is the purpose, access, and retention for images containing workers or visitors?
• May confidential customer parts or fixtures be used to train a vendor’s model?
• Who remains accountable for quality disposition, line stop, and scrap?
• Which product, camera, generator, or model changes trigger reacceptance?
• Are cybersecurity, OT connectivity, and functional safety reviewed separately from dataset quality?

Synthetic does not mean unrestricted. A generated image can reproduce confidential CAD or factory layout and still expose trade secrets. Apply asset-level information classification and contract terms.

Common implementation failures

• Counting images as the KPI: volume rises without useful diversity.
• Showing the final set to the generator team: acceptance becomes tuning.
• Judging only photorealism: human realism and model-relevant features differ.
• Ignoring hard normal samples: glare and contamination become false defects.
• Fixing one synthetic ratio: the optimum varies by defect and learning stage.
• Accepting the average: critical slices remain unsafe or uneconomic.
• Losing generator versioning: results cannot be reproduced or diagnosed.
• Stopping real collection: production drift and new failure modes go unseen.

Implementation checklist

Before the RFP

• [ ] Agree the process, defects, exclusions, and cost of errors.
• [ ] Create a real-only baseline and isolated real holdout.
• [ ] Have Quality approve ontology and label boundaries.
• [ ] Review CAD, image, and generated-asset rights.
• [ ] Build appearance, content, process, and time gap hypotheses.
• [ ] Approve stop rules and a budget cap.

During the PoC

• [ ] Manifest factors, rationale, seed, and generator version.
• [ ] Test duplicates, artifacts, labels, and confidential-content leakage.
• [ ] Compare real-only, synthetic-only, and mixed under equal conditions.
• [ ] Evaluate by defect, product, asset, light, pose, and size.
• [ ] Measure takt, review, false rejects, and downtime in shadow mode.
• [ ] Link each gap correction to its effect.

Before acceptance

• [ ] Re-evaluate on frozen real data and a later period.
• [ ] Meet critical-slice floors and process KPIs.
• [ ] Document unresolved gaps and the valid operating envelope.
• [ ] Define retest triggers for product, machine, camera, or generator changes.
• [ ] Receive data, configuration, manifest, rights, and deletion evidence.
• [ ] Rehearse rollback to the existing inspection method.

FAQ: synthetic data manufacturing deployment

Can synthetic data visual inspection eliminate all real defect images?

Normally, real data is still required for factory acceptance. Specific studies show strong synthetic-only results, but they do not prove the same transfer for your product, optics, process, or defects. Where true defects are extremely rare, combine hard normal samples, process trials, expert review, and post-launch shadow monitoring, and state the uncertainty.

Should defect image generation use 3D or generative AI?

3D is attractive when geometry, pose, occlusion, and label precision need strict control. A generative model may produce appearance candidates faster. A hybrid can use 3D for geometry and labels, generative methods for texture variation, and real images for backgrounds and sensor characteristics. Select by reproducibility, rights, gap closure, and real-data performance—not the technology label.

What percentage of real and synthetic data is best?

There is no general optimum. It depends on defect, product, model, generator quality, and fine-tuning strategy. Compare real-only, synthetic-only, and multiple mixed ratios on the same isolated real set. Use important-slice floors and process losses rather than only the aggregate.

How should the domain gap be measured?

Use image statistics and feature embeddings as diagnostics, then classify real-data errors into optical, geometric, process, sensor, semantic, and temporal gaps. The decisive measurement is which real-data slices improve or regress after adding synthetic data. Do not accept a single gap score as proof.

What are good stop criteria for a 90-day PoC?

Examples include missing the critical-defect floor, unresolved major gaps, unauditable rights or generation, excessive takt or review work, and recurring TCO above the cap. An extension should require an identified cause, defined change, deadline, cost, and reacceptance gate.

Summary: buy real-line performance and evidence, not generated images

A synthetic data manufacturing program should deliver a traceable process: scope, factor distribution, provenance, rights, generator version, labels, quality gates, real–synthetic mixing, domain-gap register, stop rules, and acceptance results. Compare real-only, synthetic-only, and mixed training fairly. Decide on an isolated real set and shadow line evidence. Research and vendor examples support hypotheses, but do not guarantee your factory outcome. Acceptance means that the important operating slices meet process requirements and that the system can be retested and rolled back after change.

TOMAS TECH can support an early “should we use synthetic data?” assessment, data inventory, RFP, 90-day PoC, and acceptance-evidence design for visual inspection and robot perception. You can contact us while you are still comparing approaches and before committing to a generation platform.

References

• NIST, “2026 Roadmap on Artificial Intelligence and Machine Learning for Smart Manufacturing” (3 July 2026): https://www.nist.gov/publications/2026-roadmap-artificial-intelligence-and-machine-learning-smart-manufacturing
• NIST, “Artificial Intelligence (AI) for Manufacturing” (updated 17 July 2026): https://www.nist.gov/programs-projects/artificial-intelligence-ai-manufacturing
• NVIDIA, “Metropolis for Factories”: https://developer.nvidia.com/metropolis-for-factories
• NVIDIA, “Omniverse Replicator”: https://docs.omniverse.nvidia.com/extensions/latest/ext_replicator.html
• Eltoum et al., “BladeSynth: A High-Quality Rendering-Based Synthetic Dataset for Aero Engine Blade Defect Inspection,” Scientific Data 12, 1268 (2025): https://www.nature.com/articles/s41597-025-05563-y
• Zhang et al., “Pseudo defect guided meta calibration for few shot industrial defect segmentation,” Scientific Reports (2026): https://www.nature.com/articles/s41598-026-63902-4
• Huber, Knoll & Guthe, “Fully-Synthetic Training for Visual Quality Inspection in Automotive Production,” Procedia CIRP 134 (2025): https://arxiv.org/abs/2503.09354
• ETDA, “AI 2026: Driving Trust AI Governance” (9 June 2026): https://www.etda.or.th/th/pr-news/aigc_Driving-Trust_AI_Governance.aspx

This article is general procurement and implementation guidance based on public information checked through 20 September 2026. It does not replace equipment-specific functional-safety assessment, quality assurance, or legal and contractual advice.