What does it take to make data AI ready?

Transitioning from FAIR to AI Ready Data in the Physical Sciences: A PSDI & AIchemy Workshop Writeup

The Transitioning from FAIR to AI-Ready Data in the Physical Sciences event (co-hosted by PSDI and AIChemy) was held at University of Southampton in July 2026. This event posed the question ‘What does it really mean for scientific data to be “AI-ready”?‘ and explored how the FAIR principles, while essential for good research data management, are not sufficient to support modern AI applications. Across presentations, lightning talks, and a working session, participants suggested that AI-ready data requires additional layers of documentation, semantic metadata, provenance, and standardisation to ensure that datasets can be understood and reused by both humans and machines. A recurring message was that AI-ready data should minimise the amount of preprocessing required by downstream users while maintaining high levels of transparency, traceability, and trustworthiness. 

Aileen Day – What did PSDI learn when making physical sciences datasets AI ready?

Aileen Day opened the event by presenting the outcomes of a six-month PSDI project aimed at understanding what is required to transform physical science datasets into AI-ready resources. The team initially followed their standard FAIR data workflow by ingesting datasets into PSDI’s Community Data Collections, defining metadata fields, exposing the records through APIs, and registering them in the PSDI Resource Catalogue. However, feedback from machine learning researchers revealed that FAIR datasets were still difficult to use for AI because users had to manually download, combine, and interpret hundreds of individual records. This led the team to redesign their approach by creating aggregated datasets packaged as a single downloadable resource accompanied by comprehensive metadata. They adopted the Croissant metadata standard to provide machine-readable descriptions of datasets, variables, provenance, licensing, and intended AI tasks, while also including RO-Crate files for research object packaging and human-readable documentation. A key lesson from the project was that AI-ready datasets require considerably richer metadata than FAIR datasets, including explicit descriptions of variables, coded values, training and validation splits, and responsible AI information. The team also distinguished between metadata that must be supplied by domain experts, such as provenance and scientific meaning, and structural metadata that can be generated automatically, highlighting automation as an important area for future work. 

Read more about the project: https://www.psdi.ac.uk/news_ai_ready_data/ 

Matthew Partridge – Making a data collection AI Ready

Matthew Partridge expanded on these ideas through the example of PChProp, PSDI’s curated physical chemistry database containing hundreds of thousands of experimentally validated measurements. He emphasised that the value of AI depends fundamentally on the quality of the underlying data, making rigorous curation an essential prerequisite for AI readiness. Before any dataset is incorporated, the team verifies its provenance, licensing, completeness, and suitability for reuse, ensuring that every measurement can be traced back to its original publication and that any future corrections or updates remain possible. While traditional datasets often consist of simple CSV files, these lack the contextual information that AI systems require. For example, an AI model cannot infer the meaning of identifiers, scientific abbreviations, units, or expected value ranges without explicit metadata. To address this, PSDI uses Croissant to describe not only datasets and files but also every variable, ontology reference, and scientific concept in a machine-readable format. The presentation also demonstrated how Python scripts can automatically generate Croissant metadata, RO-Crates, and README files, reducing the burden of preparing AI-ready datasets while maintaining consistency. Audience discussion focused on handling missing values, outliers, and contradictory measurements, with Partridge explaining that unusual values are investigated rather than discarded, and that contextual notes are recorded whenever anomalies are retained in the database. 

Nessa Carson – Building trustworthy, reusable reaction data

Nessa Carson presented AstraZeneca’s work on developing trustworthy reaction datasets for AI-assisted drug discovery and chemical synthesis planning. She described how pharmaceutical research relies on extremely large collections of chemical reactions originating from electronic laboratory notebooks, patents, scientific literature, outsourced research, and internal experimental data. These datasets contain millions of reactions but are often inconsistent, incomplete, or poorly standardised, making them difficult to reuse directly for AI. One of the major challenges is the absence of comprehensive ontologies for reaction chemistry, requiring existing vocabularies to be extended or modified to adequately represent complex chemical processes. Carson explained that while AI tools can assist in extracting information from historical documents and identifying missing information, they frequently misinterpret chemical terminology, abbreviations, and molecular structures, making expert validation indispensable. AstraZeneca therefore combines AI-assisted extraction with extensive manual curation performed by chemists, applying multiple quality-control checks before information is incorporated into their knowledge base. She also stressed the importance of capturing negative experimental results, which are rarely published despite their considerable value for training predictive models, and argued that improving AI-ready data ultimately requires closer collaboration with data generators so that richer metadata can be captured at the point of data creation rather than reconstructed retrospectively. 

Otello Roscioni – Balancing FAIRness and Data Sovereignty & Computational Materials Science

Otello Roscioni concluded the first part of the event by addressing the relationship between FAIR data and data sovereignty in computational materials science. His presentation focused on semantic technologies and federated data spaces as mechanisms for enabling collaboration while allowing institutions to retain ownership and control over their data. Using a computational chemistry case study, he demonstrated how research workflows managed through AiiDA could be connected to ontologies, knowledge graphs, and semantic web technologies to create interoperable research objects that preserve detailed provenance information. Rather than relying on centralised repositories, different organisations can maintain their own data while exposing standardised semantic descriptions that allow datasets to be discovered, queried, and linked across institutional boundaries. Through live demonstrations, Roscioni illustrated how SPARQL queries could retrieve computational workflows, simulation parameters, material compositions, and ontology definitions across multiple interconnected data spaces. This architecture enables reproducibility by exposing not only simulation outputs but also the parameters, workflows, and semantic descriptions that generated them. At the same time, it supports data sovereignty by allowing organisations to determine which parts of their knowledge graph are shared and with whom through fine-grained access controls, demonstrating that interoperability and controlled access can coexist within FAIR data infrastructures. 

Discussion sessions

The second half of the event consisted of facilitated discussion tables that explored the key considerations for transitioning from FAIR to AI-ready data. Groups of participants discussed the technical, organisational, and ethical requirements for developing AI-ready datasets, including the infrastructure, tools, standards, and skills needed to support their creation and reuse. Conversations also examined issues of sustainability, responsible AI, governance, and the balance between openness and appropriate control of data, to identify priorities and recommendations for advancing AI-ready data practices across the physical sciences.  

Across groups, the main takeaways were: 

The conversation is far from over, and we encourage those who could not attend to register for the PSDI mailing list and AIChemy mailing list to hear more about our work and events and hopefully you can join us at the next workshop and help move this important work forward. If you want to discuss these topics further feel free to reach out to our team. 

Thanks to Carolina Mayen Huerta, Digital Curation Centre, University of Edinburgh who wrote the original version of this content and Matthew Partridge, University of Southampton for additional content which were edited for website publication.