Table of content
- What “digital NDT” actually means right now
- Where AI and automated defect recognition genuinely work today
- Validation and regulation: the honest state of AI oversight in safety-critical NDT
- Data standards: DICONDE, and the open gap in AI training-data standards
- Digital twins: where the concept actually stands
- Connected instruments: a real example, correctly scoped
- Digital NDT capability, by maturity
- Common mistakes when evaluating digital NDT claims
- Key takeaways
- FAQs
- References
What “digital NDT” actually means right now
Non-destructive testing is in the middle of a marketing cycle that outruns its engineering. Vendor copy routinely stacks “AI-powered,” “digital twin,” and “NDE 4.0” onto products whose actual capability is a connected UT gauge with a cloud dashboard. That gap matters for buyers: a QA manager comparing systems needs to know which claims are validated technology in production use today, which are real but early-stage research, and which are aspirational framing with no deployed product behind them yet.
This article sorts digital NDT into four distinct, honestly scoped areas: automated defect recognition (ADR) and machine learning applied to inspection data; the regulatory and validation landscape those tools operate in (which, for safety-critical NDT, remains largely unwritten); data standards, led by DICONDE, and the open question of how AI training data should be structured and shared; and digital twins, a concept with real engineering content but very little deployed NDT-specific product. A connected-instrument example closes the piece, scoped strictly to what its vendor has actually published.
Where AI and automated defect recognition genuinely work today
ADR is the most mature of the four areas, because it has the most tractable problem shape: a trained model scores an image (a radiograph, a C-scan, a phased-array frame) against known defect signatures and flags candidates for a human reviewer. It is assistive, not autonomous, in every credible published deployment.
Weld radiography. Deep-learning models for detecting porosity, cracks, and lack-of-fusion in weld radiographs are an active, well-published research area. A 2022 peer-reviewed study by researchers affiliated with Aalto University (Tyystjärvi et al., published in Welding in the World) trained a deep-learning segmentation model on aerospace weld radiographs. It found that augmenting rare, underrepresented defect types with synthetic flaws was necessary to reach acceptable detection performance — the data-scarcity limitation recurs across the published literature. The paper does not report a direct detection-rate comparison against human readers.
Pressure-pipeline radiography. A separate 2025 peer-reviewed study (Coatings, 2025) applied deep-learning object detection to weld-defect identification in pressure-pipeline radiographs, reporting improved detection and classification accuracy over baseline methods; the published abstract does not give comparable rare-defect figures, so that specific caveat should not be assumed to carry over.
Digital radiography throughput in aerospace and automotive. Quality Magazine’s coverage of high-throughput digital radiography documents ADR software screening large volumes of castings and welds to triage parts for human review, cutting inspector time spent on clearly defect-free parts rather than replacing the final accept/reject call. The piece is authored by a product manager at North Star Imaging, the ADR vendor whose efX-ADR product supplies the cited performance figures (roughly 1.5 seconds per image versus about 60 seconds manual; 97.25% repeatability) — those figures are vendor-supplied, not independently verified, though the general assistive framing is consistent with the other sources in this section.
Ultrasonic NDT in nuclear in-service inspection. A 2023 Oak Ridge National Laboratory technical report examined machine-learning approaches for ultrasonic NDT data in nuclear in-service inspection, situating the work explicitly within the ASME Boiler and Pressure Vessel Code, Section XI inspection framework — the governing code for nuclear in-service inspection in the US. The report frames ML as a decision-support aid for data interpretation, not a code-qualified substitute for a certified Level II/III analyst’s sign-off.
The honest pattern across all four examples: every validated or credible near-term application is assistive — it screens, triages, or flags for a qualified human inspector, who retains the certifying judgment. No published source reviewed describes a deployed system in which ADR output alone, without a certified human sign-off, constitutes the accept/reject decision on safety-critical NDT.
Validation and regulation: the honest state of AI oversight in safety-critical NDT
There is currently no NDT-specific regulation or consensus standard governing how an AI/ML-based defect-detection tool must be validated before use in safety-critical inspection. ASME, API, and ASNT have not published a binding AI/ML validation standard for NDT as of this writing. Any vendor or article claiming a tool is “FDA-cleared,” “ASME-certified,” or similarly validated for AI-assisted NDT specifically should be read with that gap in mind — ask which standard, issued by whom, and covering what exact claim.
The closest real regulatory precedent, in an adjacent safety-critical field, is aviation. The European Union Aviation Safety Agency (EASA) has published an AI Roadmap and a Concept Paper on “Guidance for Level 1 & 2 Machine Learning Applications,” setting out how machine-learning components can be certified for use in aircraft systems and operations. A 2024 peer-reviewed article in Frontiers in Aerospace Engineering examines the practical challenges of certifying airborne AI against that framework.
This is presented here as an adjacent framework the NDT field is watching, not as an existing NDT rule. Aviation’s AI certification work addresses a different regulatory question — airworthiness of flight-critical software — than the question NDT eventually needs answered: how to validate that an ADR model’s detection performance on a given defect class and material is reliable enough to support an accept/reject decision under a code such as ASME Section V or Section XI. No body has yet published that NDT-specific answer. Buyers evaluating an AI-enabled inspection tool today are, in effect, doing their own validation.
Data standards: DICONDE, and the open gap in AI training-data standards
DICONDE (Digital Imaging and Communication in Nondestructive Evaluation) is the established data standard for this field, developed and maintained by ASTM International as ASTM E2339, with related standards E2663 (DICONDE for ultrasonic test methods) and E2660 (DICONDE for digital reference images). Modeled on DICOM, the medical-imaging standard it borrows its structure from, DICONDE defines how NDT image data, metadata, and examination records are formatted and exchanged so that files from different vendors’ instruments and software remain interoperable.
What DICONDE does not yet do is standardize the data an AI model needs to train on: labeled defect datasets, annotation formats for ground-truth defect locations and classifications, or a shared structure for digital-twin data. No ASTM, ASME, or ASNT standard committee has published a distinct standard for AI-training-data or labeled-defect-dataset structuring in NDT as of this writing. This is a genuine, nameable gap rather than an oversight on the part of any one vendor: every ADR research effort cited in this article (Aalto, the pressure-pipeline study, ORNL) assembled and labeled its own training data, because no shared, standardized labeled dataset or annotation format yet exists for the field to draw on. Until that changes, AI training-data quality and provenance remain something a buyer must evaluate vendor by vendor, with no standard to check against.
Digital twins: where the concept actually stands
“Digital twin” is used in NDT discussions to describe a live, data-linked virtual model of a physical asset — one that absorbs inspection results over time and is used to project remaining life, plan future inspection intervals, or simulate how a known defect might progress. The concept has real engineering content in adjacent fields: aerospace structural health monitoring and nuclear plant asset management both have published digital-twin research programs that explicitly incorporate inspection data as an input.
For NDT specifically, deployed, commercially available digital-twin products are limited. Most of what is marketed as a “digital twin” in NDT vendor material today is, on inspection of the underlying product, closer to a connected data dashboard: inspection history, instrument readings, and asset metadata aggregated in one place, without the predictive simulation layer that distinguishes a true digital twin from a historical database. That distinction matters for buyers evaluating a “digital twin” claim: ask specifically whether the product simulates future defect progression or asset condition, or whether it aggregates and displays past inspection data — the second is valuable and real, but it is not a digital twin in the engineering sense of the term.
The connected-instrument example in the next section illustrates the data-aggregation layer that digital twins for NDT would need to build on, without itself claiming to be a digital twin.
Connected instruments: a real example, correctly scoped
Pragma NDT, a Quebec-based instrument maker, illustrates the connectivity and data-fusion layer that digital NDT is actually built on today. In a July 2024 interview in LEMO’s Connected magazine, founder and CEO François Mainguy described the company’s PragmaFlex platform’s approach to combining ultrasonic, phased-array ultrasonic, and eddy current data into a single connected workflow. A June 2025 BINDT item covering PragmaFlex’s platform/product coverage listed eddy current array, bond testing, and microtopography among its supported techniques.
In June 2024, Pragma NDT and GelSight announced a partnership pairing Pragma’s platform with GelSight’s 3D surface-topography sensing. GelSight CEO Youssef Benmokhtar framed the partnership around fully automated NDT workflows and finer-grained, micrometer-level measurement data; Mainguy’s own comments in the release focused on the data-fusion and connectivity value of bringing another modality into PragmaFlex, calling it a natural extension of the platform.
None of this published Pragma material mentions AI or machine learning. It is cited here strictly as a real, verifiable example of multi-modality data fusion and connected-instrument workflows — the NDE 4.0 connectivity layer discussed above — not as an AI or digital-twin example. Characterizing Pragma as an AI example would overclaim beyond what the company has actually published.
Digital NDT capability, by maturity
Capability | Maturity today | Real example cited above | Caution for buyers |
|---|---|---|---|
ADR for weld radiography (porosity, cracks, lack-of-fusion) | Validated research, assistive deployment | Aalto University study; pressure-pipeline study | Rare defect classes still underperform due to limited training data |
ADR for high-throughput DR triage (castings, welds) | Deployed, assistive | Quality Magazine coverage | Screens for human review; does not replace the final accept/reject call |
ML-assisted UT interpretation in nuclear in-service inspection | Early deployment, code-adjacent | ORNL 2023 report (ASME Section XI context) | Decision-support only; certified Level II/III sign-off still required |
AI/ML validation standard for NDT | Does not yet exist | None — adjacent precedent only (EASA aviation AI framework) | No binding NDT standard to check a vendor’s claim against |
DICONDE data interoperability | Established standard | ASTM E2339 / E2663 / E2660 | Covers image/metadata exchange, not AI training-data labeling |
AI training-data / labeled-defect-dataset standard | Open gap, no standard | None published | Every cited study built its own training set; no shared benchmark exists |
True predictive digital twin for NDT assets | Mostly aspirational marketing | Adjacent fields only (aerospace SHM, nuclear asset management) | Most “digital twin” NDT products are connected dashboards, not simulators |
Connected multi-modality instrument platforms | Deployed | Pragma NDT / PragmaFlex | Real and useful, but a connectivity layer, not an AI or twin capability |
Common mistakes when evaluating digital NDT claims
- Treating “AI-powered” as a validated capability claim. No NDT-specific AI/ML validation standard exists yet; the phrase describes a technology choice, not a certification.
- Accepting “digital twin” at face value. Ask whether the product simulates future condition or defect progression, or whether it displays past inspection data in a dashboard. Most commercial NDT “digital twin” products today are the latter.
- Assuming ADR replaces the certified inspector’s sign-off. Every credible published deployment reviewed is assistive — screening or triaging for human review, not an autonomous accept/reject decision.
- Confusing DICONDE with an AI data standard. DICONDE (ASTM E2339 and related standards) governs image and metadata interoperability. It does not standardize labeled training data or annotation formats for machine learning.
- Taking a vendor’s modality or performance claims from marketing copy alone. Where this article cites a vendor capability (Pragma NDT, for example), it traces the claim to a specific published interview, magazine item, or press release — not a product page. Ask vendors for the same.
- Assuming aviation’s AI certification framework already applies to NDT. EASA’s AI Roadmap is a genuine, citable precedent in an adjacent field, but it has not been adopted as an NDT standard by ASME, API, or ASNT.
Key takeaways
- Real, validated AI/ADR applications in NDT exist today, but every credible example is assistive — screening or triaging data for a certified human inspector’s sign-off, not replacing it.
- No NDT-specific regulatory body (ASME, API, ASNT) has published a binding AI/ML validation standard. The closest precedent is EASA’s aviation AI certification framework, which is adjacent, not adopted, for NDT.
- DICONDE (ASTM E2339 and related standards) is the established data-interoperability standard for NDT, but it does not cover AI training-data or labeled-defect-dataset standardization — a genuine open gap.
- “Digital twin” is mostly aspirational marketing in NDT today; most products sold under that name are connected inspection-history dashboards, not predictive simulators.
- Vendor claims should be traced to specific published, dated sources, not product-page copy. Pragma NDT’s real published material supports a connectivity/data-fusion example, but does not support an AI claim — a distinction worth checking for any vendor.
FAQs
Is AI allowed to make the accept/reject decision in safety-critical NDT? Not under any current code or standard. Every validated published deployment uses AI/ML as a screening or triage aid, with a certified human inspector (Level II/III, depending on the code) retaining the sign-off.
Is there a certification for AI-based defect-detection software in NDT? No NDT-specific one exists yet from ASME, API, or ASNT. Treat any claim of “certified AI” for NDT with that gap in mind, and ask the vendor exactly what was certified, by whom, and against what standard.
What is DICONDE, and does it cover AI data? DICONDE (ASTM E2339, with related standards E2663 and E2660) is the established standard for formatting and exchanging NDT image and metadata, modeled on the medical DICOM standard. It does not cover labeled training-data or annotation formats for machine learning — that remains unstandardized.
Are digital twins actually used in NDT today? Rarely, in the full engineering sense of a predictive simulation model. Most NDT products marketed as “digital twins” are connected dashboards aggregating historical inspection data, which is valuable but not the same thing as simulating future defect progression.
Is Pragma NDT an AI company? Pragma NDT’s published material (LEMO Connected, BINDT, and its GelSight partnership announcement) describes multi-modality data fusion and connected-instrument workflows. None of that published material mentions AI or machine learning, so this article does not characterize Pragma as an AI example.
References
- Deep Learning for Weld Defect Detection in Aerospace Radiographs – Aalto University, 2023
- Deep-learning object detection for weld-defect identification in pressure-pipeline radiographs — peer-reviewed study, Coatings, 2025.
- “High-Throughput DR: Aerospace and Automotive Inspection with ADR” – Quality Magazine
- Machine Learning for Ultrasonic NDT in Nuclear In-Service Inspection – Oak Ridge National Laboratory, 2023 (DOI 10.2172/2320378)
- EASA Artificial Intelligence Roadmap
- EASA Concept Paper: Guidance for Level 1 & 2 Machine Learning Applications
- Certifying airborne AI: a 2024 review – Frontiers in Aerospace Engineering.
- ASTM E2339: Standard Practice for Digital Imaging and Communication in Nondestructive Evaluation (DICONDE)
- LEMO Connected magazine, “Pragma NDT” feature, 8 July 2024.
- BINDT, Pragma NDT platform update coverage, June 2025.
- GelSight / Pragma NDT partnership announcement, June 2024.




There are no comments yet!