Reliable AI in Material Science: A FAIR-by-Design Path from Data to Services

Materials science faces a dual challenge: transforming legacy archives into FAIR-compliant datasets through retrospective curation (FAIRification), then establishing prospective workflows that embed FAIR principles from inception (FAIR-by-design). This work illustrates a FAIR-by-design path from curated data to deployed AI services, addressing challenges in experimental microscopy and spectroscopy. A FAIRification foundation was established by curating a legacy Scanning Tunneling Microscopy (STM) archive into public datasets with rich metadata and formal provenance. Building on this, the research developed a suite of artificial intelligence (AI) models to enhance experimental data with methods for STM artifact detection and generative restoration, and a dual framework for Near-Edge X-ray Absorption Fine Structure (NEXAFS) signal decomposition based on deep learning and Bayesian approaches. The research concludes with the deployment of these models as operational open-access services within existing European nanoscience infrastructure. Collectively, the contributions of this thesis provide a reproducible methodology that connects principled data stewardship to the creation of reliable, deployable AI tools for the scientific community.