Turning Collections Into Data: Unlocking Hidden Stories at the Science Museum Group

What happens when museums stop treating their collections as static objects, and instead see them as data?

That’s the question Daniel Belteki, Digital Research Applications Fellow at the Science Museum, has been exploring. Speaking from the National Railway Museum’s Media Store, Belteki explained that “these experiments are about more than just technology. They are about what it takes to make collections more discoverable, meaningful and connected in a digital world.”

The Science Museum Group’s approach can be summed up in three explorations: mapping places, linking people, and tracing networks from a single object. Each case study reveals not only the promise of digital tools, but also their limitations.

Mapping the Geography of Collections

The Science Museum Group’s collections hold around 10,000 location entries. These range from nations and cities to demolished buildings and railway stations. Historically, these entries existed in isolation—useful in detail, but inert in digital form.

“We asked, how do we liberate these places? How do we make them useful? Discoverable beyond the database?” said Belteki.

The answer came through linking their location data with the Getty Thesaurus of Geographic Names. With OpenRefine, the team reconciled their place entries to Getty, creating spatial visualisations that revealed regional emphases, absences, and unexpected connections.

But this experiment also revealed a sobering truth. “If a place isn’t represented in structured data, it disappears,” Belteki explained. “It’s as though the object was never there. This is the paradox of digital representation. No data, no presence, no digital liberation.”

Railway posters became a fascinating testing ground. Using Google’s Gemini large language model, the team extracted destinations hidden in catalogue descriptions and matched them to their cleaned dataset. The result was a map not just of where trains went, but of where metadata hadn’t yet travelled.

As Oliver Betts, Research Lead at the National Railway Museum, put it: “For us, trying to find those [locations]… that information’s buried in a descriptor field on our catalogue. So it makes it really hard to find that location as a data point for research and for finding aids.”

Betts believes the posters are more than travel advertising—they’re part of Britain’s art history and social memory. “Thinking about how these posters map onto the British seaside right at the height of the railways at the end of the 19th and early 20th century… but now after decades of closures where those posters and locations have ended up stranded from the rail network, that would be a fascinating thing to explore.”

Linking People Across the Collection

If places bring structure, people bring narrative. The Science Museum Group holds around 130,000 person and organisation records. Yet most are minimal: just a name tied to an object.

Here too, digital tools offered a way forward. The team applied large language models not just to compare entries, but to reason through ambiguities—whether, for instance, an optician named William Harris was the same man as a mathematical instrument maker.

“The combination of existing programming tools enhanced with the LLM and other AI applications greatly increases the scale at which we can operate,” explained Lawrence Brooks, the Group Collections Data Manager. “The LLM brings in more of what was once a human task—the reasoning or the reasoned response—that otherwise a human would’ve done but possibly wouldn’t have recorded.”

For Brooks, the future lies in linked open data: “not a series of discrete bastions of shared knowledge, but an interrelated web of shared knowledge whereby no one place has to be the authority.”

From One Object to a Constellation: Hawking’s Blackboard

Sometimes the best way to think about collections as data is to start not with the dataset, but with a single object. For the Science Museum Group, that object was Stephen Hawking’s blackboard, inscribed with equations, doodles, and jokes during a 1980 Cambridge workshop.

“It’s one of the best examples we have of how collaborative science is in the late 20th century,” said Juan-Andres Leon, Curator of Physics at the Science Museum. “The blackboard was done actually during a scientific conference… attendees would just start drawing basically graffiti on it as the conference went along. So at the end, you had this record of what people were thinking throughout the conference and who was connected to whom.”

The team used a combination of Gemini, ChatGPT, Python, Regex, and OpenRefine to extract and reconcile names from the blackboard and related archives. The result: a network map linking attendees to books, objects, and collections across institutions.

Leon sees enormous potential in this approach. “Because it’s a scientific conference, there are proceedings listing all the attendees. One of the obvious questions is how do the cartoons relate to the people who attended? Who was close to Hawking’s research group, and who was just attending as a one-off? If there’s a tool that can systematically extract that information and relate it to what’s elsewhere, that creates a whole new scale of possibilities.”

Beyond Technology: The Questions That Remain

Across these case studies, a theme emerges. Digital tools are not neutral. They can connect scattered histories, reveal gaps, and generate new meaning. But they also bring ethical, legal, and technical complications.

As Belteki concluded: “They remind us, data is not neutral. It’s shaped by what we choose to record and what we don’t. It’s shaped by how we process them and how we deploy them. So perhaps the final deliberation is this, not whether to use digital tools, but how to use them.”

For museum professionals, the lesson is clear: data is more than metadata. It is storytelling. And how we structure, connect, and interpret it determines what stories survive into the digital future.

–

Daniel Belteki, Digital Research Applications Fellow at the Science Museum spoke at the Digital Collections Summit 2025.