Horizon Europe · HORIZON-INFRA-2025-01-DEV-03 · M1–M36
AI4Access turns the Euro-BioImaging catalogue into a machine-actionable knowledge base — an ontology-driven graph of services, methods, workflows and providers, linked by persistent identifiers. That graph is the project. This is a working reference to it, read from the seat that builds it — unfamiliar terms are in the glossary.
Where we sit
Entities, relations, controlled vocabularies, PID strategy. What counts as a service, how a facility relates to a technology, which term is authoritative. Decided in T2.1, inherited by everything after it.
39 Nodes, a website full of prose, and 1,100+ past projects — collected, enriched with AI assistance, reviewed by humans, and kept fresh by an update protocol that outlives the grant.
KTH's backend, the MCP SDK, the pilots, the provenance framework, the explanations shown to users — all read from what we build. When the schema slips, seventeen tasks slip with it.
12 of 33 tasks led or co-led here
The foundation stone. Define the minimum viable semantic model for services, technologies, facilities and workflows; pick and align controlled vocabularies; set the PID strategy (ORCID / ROR / DOI / PIDINST) that every later record hangs off.
Where the model meets 39 Nodes and 295 facilities. Templates go out, messy reality comes back. AI-assisted enrichment plus human review loops turn heterogeneous Node descriptions into the first harmonised dataset.
The public website and Access Portal already hold years of curated knowledge in prose. Map it, tag it semantically, parse it with AI assistance, and turn it into machine-actionable content the Navigator can cite.
A knowledge base that is not maintained is a knowledge base that lies. Design the update cycle, self-service editing, AI-assisted quality checks and the governance that keeps it alive after M36.
Cross-RI semantic mapping. INFRA-SERV catalogue vocabularies (canSERV, ISIDORe, AgroSERV) will not match ours; this task owns the crosswalks and proves that a query can route across infrastructure boundaries without losing meaning.
Survey the analysis-tool landscape and score integration maturity and FAIRness. BIII is the obvious anchor: reuse its tool/workflow records rather than re-cataloguing from scratch.
The longest-running task led here — M1 to M36. Maintain the machine-actionable image-data service catalogue and drive EDAM / EDAM-Bioimaging updates back to the upstream community, not just into this project's own store.
Provenance, versioning and rollback are data-architecture problems wearing an engineering hat. The T2.4 update protocol and this pipeline must be the same mechanism, not two competing ones.
Five pilots test whether the model survives contact with real national information systems. Database integration here is a mapping exercise against local schemas this function does not control.
Facility management systems (booking, LIMS) hold the operational truth. Mapping them defines the data-sharing boundary between what the Navigator may know and what stays local.
The update protocol (T2.4) and the SDK/MCP adapters (T4.4) are only as good as whether a Node coordinator can actually operate them. This task turns both into workshops, checklists and an adoption playbook.
"Why did the Navigator recommend this?" is answerable only if the evidence objects and their links were modelled from day one. Model/data cards and audit logs draw straight from the provenance graph built here.
The spine
7 open items, 2 critical
T2.2 asks 39 Nodes to submit structured service descriptions. Without a real intake surface — templates, validation on entry, per-Node status, review queue — collection degrades into spreadsheets by email, and the harmonisation cost lands here at exactly the moment D2.2 (M18) is due.
AskAgree by ~M4 whether this is built in-house, extended from the Access Portal, or bought. It has to exist before the model is finalised, because the intake form is the model made visible.
T2.1 must pick between extending EDAM-Bioimaging, formalising the 2022 harmonised services ontology, adopting foundingGIDE recommendations, or a hybrid. Every downstream annotation, crosswalk and adapter schema inherits this choice, and reversing it after M12 is expensive.
AskA time-boxed decision workshop before M6, with a written rationale. Recommendation: OLS-resolved terms, EDAM-Bioimaging for analysis, a small Euro-BioImaging services extension for access concepts.
T4.1 delivers the backend and hybrid knowledge base at M18 and depends on T2.1–T2.4. If the model is still moving at M15, KTH builds against a shifting target and both sides absorb rework.
AskAgree a v1.0 schema freeze at M12 with a versioned change process afterwards — additive changes only until M18.
The register lists 'new hire' on T1.1–T1.4 and 'Aman till new hire' on WP4 tasks. Inputs from this function land on people who are not yet in post, and the WP1 persona work that shapes the data model is among them.
AskConfirm start dates so it's known whether the T1.2 intent taxonomy will really be usable before the T2.1 freeze.
Node submissions (T2.2), website content (T2.3) and facility systems (T5.2) will disagree about the same service. Without a precedence rule and a sign-off owner, the knowledge base contains contradictions the Navigator will confidently repeat.
AskDefine a source precedence order and a named authority per record type, inside the T2.4 update protocol.
10 sources we build on
Proposal form fields → target schema for the agentic pre-population in T4.3
Metadata exchange schema patterns — do not reinvent the interoperability layer
Term resolution service — resolve and validate every vocabulary term at ingest, no local copies of ontologies
Operations and topics vocabulary for describing analysis services
The eight-module structure as a template for what a Node service record needs to carry
Tool and workflow records as ready-made seed content for T3.1 and T3.2
Precedent architecture for WP4's AI Agent: orchestrating heterogeneous tools and data behind one agent interface
The service-oriented portfolio structure as the skeleton of the T2.1 model
ROR for Nodes, facilities and partner organisations
Their catalogue schemas as crosswalk targets, defined once and maintained
Four organisations
Coordinator; Access Portal, personas, regulatory, INFRA-SERV, dissemination.
Home of the WP2 data model, WP3 catalogues, Node engagement.
Co-lead on WP2/WP3/WP5; medical imaging perspective on the model.
GenAI backend, hybrid knowledge base, RAG/CAG, provenance, evaluation.
The end goal
Every part of that sentence is a data-architecture commitment. "Real technology" means a maintained, ontology-aligned catalogue. "Real facility" means PID-linked Node records that are still true today. "With citations" means provenance modelled into every assertion from M1. If the graph is right, the Navigator is useful. If it is not, no amount of model quality rescues it.
8 deliverables in the register carry our name. The rest of the project reads from what they produce.