The INTERSECT Scientific Data Layer�

August 24, 2026, Pisa, Italy

Sheik Ghafoor

Tennessee Tech University

Standards-compositional provenance for automated science workflows

Swen Boehm^1,2 , Craig A. Bridges¹ , Patrick Widener¹ , Terry Jones¹ , Sheikh Ghafoor² , Christian Engelmann¹ , and Olga Kuchar¹

¹ Oak Ridge National Laboratory

² Tennessee Tech University

ORNL IS MANAGED BY UT-BATTELLE LLC �FOR THE US DEPARTMENT OF ENERGY

1

Disclaimer

​

The views expressed in this presentation are of the authors and presenters not official communications from Oak Ridge National Laboratories or US Department of Energy

2

From the scientific method to a data model�

GROUNDING THE FRAMEWORK IN HOW SCIENCE IS DONE

3

Example1: Autonomous Chemistry Laboratory (ACL)

4

ACL context: where lineage complexity appears�

Physical lineage: each processing

step creates or transforms a material

artifact.

Digital lineage: parameters, logs,

data, code, and provenance remain

connected.

Iteration: pelletization, sintering,

cooling, and grinding may repeat.

5

Spallation Neutron Source

6

Example2: Quantum Science workflow�link experimental evidence with simulation�

Experimental lineage: sample synthesis

and measurement connect material

to data.

Simulation lineage: model and code

parameters and run records create

predictions.

Convergence: measured and simulated

results meet to refine the model.

7

Scientific workflow showing physical and digital assets produced and consumed at each step.

8

Why this is hard in autonomous laboratories�

2 artifact classes: physical samples and digital

outputs

Both must remain linked across transfers, transforms, and

publication.

N heterogeneous systems: instrument APIs,

formats, storage, and timing models

Without a common semantic layer, provenance breaks at

handoffs.

9

Lineage as a query, not a document�

# from a published file back to its precursor

SELECT ?activity ?sample ?precursor WHERE {

acl:xrdDist789 prov:wasDerivedFrom ?result .

?obs sosa:hasResult ?result ;

prov:wasGeneratedBy ?activity ;

sosa:hasFeatureOfInterest ?sample .

?sample prov:wasDerivedFrom ?precursor .

}

What the audience should take away

The full chain is machine-traversable in a

single query.

Discovery and audit reuse the same graph

and the same standards.

No bespoke lineage format is required.

10

Objective

Give complex scientific workflows a data model where every artifact is described, linked, and traceable by construction.

​

Model the artifacts

Represent both digital and physical artifacts of

a workflow in one machine-interpretable

model.

Preserve lineage

Capture unbroken provenance for those

artifacts, from precursor to published result.

Separate metadata from storage

Keep semantic metadata independent from

artifact bytes, linked through stable URIs.

Stay composable

Build on established standards so domains can

extend the model without forking it.

Design goal: make data FAIR by construction, captured during workflow execution rather than

curated after the fact.

11

Related Work�

FAIR and workflows

Wilkinson et al. (2016). The FAIR Guiding Principles for scientific

data management and stewardship.

Goble et al. (2020). FAIR Computational Workflows.

Jacobsen et al. (2020). FAIR Principles: Interpretations and

Implementation Considerations.

Wilkinson et al. (2025). Applying the FAIR Principles to

Computational Workflows.

Data work and provenance

Sambasivan et al. (2021). Everyone wants to do the model work,

not the data work.

Hecht et al. (2018). Minimum Information about Engineered

Organisms (MIEO).

Moreau et al. (2013). PROV-DM: The PROV Data Model.

Bai et al. (2024). A Dynamic Knowledge Graph for Distributed Self-

Driving Laboratories.

12

What is new in this work�

Composition over

monolith

Use established W3C

ontologies for separate

concerns rather than one

custom schema.

First-class sample

lineage

Physical sample identity and

derivation are explicit

alongside digital provenance.

FAIR by construction

Discovery, access, and

provenance are operational

behaviors of services, not only

documentation.

Positioning: reusable standards-based data

layer for operational labs, complementary to

PID-graph publication ecosystems.

Curation by design: Schemas, identifiers,

validation, and provenance are planned with

domain curators and captured during workflow

execution, reducing post hoc cleaning and

metadata reconstruction.

13

Semantic composition model�

SSN/SOSA

systems and sensors

observations and results

features of interest and

samples

PROV-O

entities, activities, agents

attribution and

responsibility

end-to-end lineage

chains

DCAT

dataset and distribution

metadata

catalog discoverability

data-service exposure

Key design rule: resources can carry multiple types, so one graph remains jointly queryable while

each concern evolves independently.

14

End-to-end provenance trace from precursor to cataloged distribution

acl : precursorMixture456 a prov:Entity, sosa:Sample .

acl : synthesisRun123 a prov : Activity ;

prov : used acl : precursorMixture456 ;

prov : generated acl : sample789 ;

prov : startedAtTime "2025 -07 -10 T09 :00:00 Z "^^ xsd : dateTime ;

prov : endedAtTime "2025 -07 -10 T11 :30:00 Z "^^ xsd : dateTime .

acl : sample789 a prov : Entity , sosa : Sample ;

prov : wasGeneratedBy acl : synthesisRun123 ;

prov : wasDerivedFrom acl : precursorMixture456 .

acl : analysisBatch890 a prov : Activity ;

prov : used acl : sample789 ;

prov : generated acl : xrdObs123 .

acl : xrdObs123 a sosa : Observation , prov : Entity ;

sosa : hasFe atureOf Intere st acl : sample789 ;

sosa : observedProperty acl : phaseComposition ;

sosa : madeBySensor acl : XRDSystem ;

sosa : hasResult acl : phaseCompResult123 ;

prov : wasGeneratedBy acl : analysisBatch890 .

acl : xrdDist789 a dcat : Distribution , prov : Entity ;

prov : wasDerivedFrom acl : phaseCompResult123 ;

dcat : downloadURL < https://data.example.org/minio/acl-xrd/scan789.xrdml> .

acl : xrdDataset a dcat : Dataset ;

dcat : distribution acl : xrdDist789 .

15

DCAT example representing Data Services, Datasets and Distributions

acl : xrdRawService a dcat : DataService ;

dct : title " XRD Raw Data Store " ;

dcat : endpointURL <https://data.example.org/minio/acl-xrd-xrdml/> ;

dcat : servesDataset acl : xrdDataset ;

dct : conformsTo < http://www.w3c.org/ns/dcat> ;

dcat : contactPoint acl : aclDataManager .

acl : xrdDistXRDML a dcat : Distribution ;

dct : title " XRD Raw Data ( XRDML Format ) " ;

dcat : downloadURL <https://data.example.org/minio/acl-xrd-xrdml /scan789.xrdml> ;

dct : format "application/x-xrdml" ;

dcat : mediaType “application/xml" ;

dcat : accessService acl : xrdRawService ;

dct : issued "2025-07-10 T12 :15:00 Z "^^ xsd:dateTime ;

dcat : byteSize 2048000 .

16

Compositional typing: one resource, three roles�

# one result participates in all three concerns

acl:xrdObs123 a sosa:Observation, prov:Entity ;

sosa:hasFeatureOfInterest acl:sample789 ;

sosa:madeBySensor acl:XRDSystem ;

sosa:hasResult acl:phaseCompResult123 ;

prov:wasGeneratedBy acl:analysisBatch890 .

acl:xrdDist789 a dcat:Distribution, prov:Entity

;

prov:wasDerivedFrom acl:phaseCompResult123 ;

dcat:downloadURL <.../scan789.xrdml> .

Why it matters

No monolithic schema; each ontology

stays authoritative for its concern.

Multiple

rdf:type

declarations let one

node bridge observation, lineage, and

publication.

Concerns evolve independently while

remaining jointly queryable.

17

Service architecture: from acquisition to discovery�

Operational flow: Registry mints URI -> Repository writes named graph -> Storage persists artifact -

> Catalog indexes distribution for discovery.

18

End-to-end campaign trace�

Precursor

prov:Entity

Synthesis

prov:Activity

Sample

sosa:Sample

Observation

sosa:Observation

Distribution

dcat:Distribution

What this enables

Direct lineage query from published file to

precursor entity

Responsibility tracking across human and

software agents

Reproducibility audits across campaigns

19

How the traces are generated�

PROVENANCE IS CAPTURED AT INGESTION, NOT CURATED LATER

Instrument event

control software signals acquisition complete

Registry Service

creates a persistent URI for the new resource

Repository Service

writes a named graph with SSN/SOSA + PROV triples

​

Storage Service

persists the raw artifact (MinIO) as dcat:Distribution

​

Catalog Service

indexes the distribution for discovery

Result

linked, queryable lineage with no manual step

Implementation notes

• FastAPI + RDFLib microservices; shared Python SDK for LDP and RDF serialization.
• Platform definition (systems, procedures, properties) loaded as Turtle at deployment.
• Each acquisition maps to one named graph, keyed by the minted URI.

Why this design

• Provenance edges are written as a side effect of normal ingestion.
• Metadata graph and artifact storage stay separate but linked by stable URIs.
• Named graphs give per-resource lifecycle and isolation.

20

Lineage as a query, not a document�

# from a published file back to its precursor

SELECT ?activity ?sample ?precursor WHERE {

acl:xrdDist789 prov:wasDerivedFrom ?result .

?obs sosa:hasResult ?result ;

prov:wasGeneratedBy ?activity ;

sosa:hasFeatureOfInterest ?sample .

?sample prov:wasDerivedFrom ?precursor .

}

What the audience should take away

The full chain is machine-traversable in a

single query.

Discovery and audit reuse the same graph

and the same standards.

No bespoke lineage format is required.

21

Mapping of ACL workflow elements to ontological representations�

22

FAIR mapping as concrete system behavior�

Findable

Persistent URI assignment and

catalog indexing support

searchable discovery.

Accessible

LDP and HTTP interfaces with

content negotiation expose

metadata and linked

resources.

Interoperable / Reusable

W3C vocabularies plus full

PROV chain preserve

machine-interpretable

context.

PID strategy: operational URIs for high-frequency ingestion, mapped to DOI/Handle/ARK at

publication boundaries.

23

FAIR principle alignment

24

Metadata and artifacts, cleanly separated�

Metadata layer

RDF named graphs hold semantic context

and provenance.

Exposed through LDP containers and a

SPARQL endpoint.

Persists independently of the underlying

files.

Artifact layer

Raw and processed files live in object

storage.

Linked from metadata via

dcat:Distribution

and stable URIs.

Storage backends can change without

breaking lineage.

Payoff: metadata and data scale independently, and the semantic graph remains the durable source

of truth.

25

Current and future evaluation�

What is demonstrated now

Representational coverage for full ACL

workflow scope

Concrete FAIR mechanism mapping to

service behavior

Worked provenance trace across physical

and digital chain

What is deferred

Ingestion throughput under load

Median and p95 query latency

Graph growth and failover recovery

behavior

Message to audience: this is prototype-stage software with a feature-complete semantic model for

the ACL scope.

26

Transferability beyond ACL�

Porting recipe

1. Define domain observable properties and instrument catalog entries

2. Bind local workflow activities and agents in PROV terms

3. Register datasets, distributions, and services in DCAT

Changes required

Domain term bindings and service integration

points.

Stable core

SSN/SOSA + PROV-O + DCAT composition and

lifecycle pattern.

27

Limits, roadmap, and collaboration invite�

Current boundary

Prototype deployment maturity, with qualitative

validation in ACL campaign scope.

Next phase

Cross-lab pilots and controlled benchmark

corpus for throughput, latency, growth, and

robustness.

Takeaway: provenance quality should be engineered as infrastructure. Standards composition

offers a practical route to FAIR-by-construction scientific workflows.

28

Takeaways and call to action�

Three things to remember

Provenance quality is infrastructure, not

documentation overhead.

Composed standards reduce lock-in and

improve portability.

FAIR-by-construction is achievable in

operational automated labs.

29

Conclusion

• How would we do this without SDM frameworks
• The SDL produces high-quality data by design

​

Where we go next

Cross-lab pilots beyond the ACL.

A shared benchmark corpus for

throughput and latency.

Failover and robustness characterization

under realistic load.

30

Acknowledgements

This manuscript has been authored by UT-Battelle, LLC under Contract No. DE-AC05-00OR22725 with the U.S. Department of Energy. The United States Government retains and the publisher, by accepting the article for publication, acknowledges that the United States Government retains a non-exclusive, paid-up, irrevocable, world-wide license to publish or reproduce the published form of this manuscript, or allow others to do so, for United States Government purposes. The Department of Energy will provide public access to these results of federally sponsored research in accordance with the DOE Public Access Plan (http://energy.gov/downloads/doe-public-access-plan).

Research sponsored by the Laboratory Directed Research and Development Program's INTERSECT Initiative of Oak Ridge National Laboratory.

​

31

Questions

32