TDCC-NES · Natural & Engineering Sciences
Bridging the distance between data producer and data reuser through automated metadata generation.
Datasets are documented once, for an audience their creators can only guess at. BridgeMD uses large language models to generate documentation written for a specific group of reusers — built into the repository software researchers already use.
Project snapshot
Data producers cannot anticipate every future reuser, and documenting a dataset for all of them is prohibitively costly. So datasets stay underused outside the community that made them.
In construction materials research alone, materials scientists, process engineers and civil engineers generate different data, use different methods and different terminology — while each would benefit from reusing the others' data.
The same gap runs between algorithm researchers and domain specialists in machine learning, and across the omics datasets that plant biosynthesis research needs to combine.
Methods that turn a dataset, its existing metadata and a reuser profile into contextualised descriptions, vocabulary mappings and example code.
The open-source system built into OpenML and a second repository stack, so it reaches researchers where their data already lives.
Quality metrics and a shared evaluation protocol that test AI-generated metadata for reliability across three very different domains.
Work packages
Two build the system, three test it in a demanding domain, one takes it to the wider community.
Prompt engineering and fine-tuning of LLMs to turn a dataset, its metadata and a reuser profile into tailored documentation — plus the evaluation protocol that tests whether it is reliable.
Builds the system into OpenML and one further repository stack, through each project's own open-source community process, so the methods outlive the project.
Generates and evaluates metadata for OpenML datasets so they can be picked up in machine learning pipelines by scientists from outside the field.
Tests whether generated metadata can bridge materials scientists, process engineers and civil engineers — a direct test of the reuser profiles from WP1.
Generates metadata that lets transcriptomics and metabolomics datasets be paired via the common ontologies already used in the field.
Runs two open datathons where participants bring their own data, coordinates the consortium and carries the results to the wider research data community.
Evaluation use cases
Each use case is led by an expert in the field and applies the same two-step test: generate metadata with the WP1 methods, then evaluate its quality against a protocol built with WP1.
OpenML holds thousands of richly described datasets, yet much of that documentation is not written for the scientists who want to reuse them. A platform and a test case at once.
A sector of 18 million people in the EU, with a value chain running from raw materials to maintenance. RILEM's technical committee is building the metadata practice this use case plugs into.
Pairing transcriptomics with metabolomics reveals biosynthesis pathways, but only where metadata is aligned across omics resources — today a resource-intensive manual job.
Community testing
Participants bring their own data and put the system to work on it. Four organisations have already committed, each bringing a different kind of test.
SURF
Metadata generated from compute artifacts during active analysis, where provenance is usually lost.
Utrecht University
The university repository, covering veterinary sciences, humanities and social sciences.
TU Eindhoven
The TU/e Research Cockpit, a central service for navigating research data management.
UvA Library
The central RDM repository of the university, through the local Digital Competence Centre.
Partners
Academic partners
Community affiliates & co-funders
The datathons are open to all interested parties. Get in touch with Dr. Daphne Miedema at d.e.miedema@uva.nl.