An entire Herculaneum scroll has been read from beginning to end for the first time in history, using machine learning and computer vision — the physical document never touched, never unrolled, never at risk of crumbling to ash.

The milestone comes via the Vesuvius Challenge, an open research competition that has been crowdsourcing exactly this problem: recovering legible text from carbonized papyrus scrolls buried under volcanic mud and ash when Mount Vesuvius erupted in 79 AD. The library at Herculaneum — the only intact library to survive from antiquity — was sealed by that eruption and has remained largely unreadable ever since, despite the scrolls' physical recovery centuries ago. Attempts to unroll them mechanically destroyed many; the remainder sat as fragile, opaque carbon cylinders.

The Vesuvius Challenge 2023 Grand Prize was awarded to a team that successfully extracted readable passages from an unopened scroll — 15 passages comprising more than 2,000 characters, decoded in February 2024. That result was already a proof of concept. The full-scroll read announced via the challenge's site extends that work to the complete document: start to finish, intact.

How the Pipeline Works

The core technique chains several well-established components in a non-obvious way. Micro-CT scanning — the same class of X-ray tomography used in industrial non-destructive testing — captures volumetric density data from a sealed scroll at sub-millimeter resolution. From that 3D dataset, researchers computationally "unwrap" the scroll's layers, essentially reconstructing the geometry of a rolled manuscript without touching it. The result is a flattened virtual surface that can be analyzed for ink traces.

The ink detection step is where ML becomes indispensable. Carbon-based ink on carbonized papyrus produces almost no contrast in a standard CT scan — both substrate and inscription are carbon. The Vesuvius Challenge teams trained models to detect the faint crackle-texture signatures that ancient ink leaves on papyrus at a microscopic level, a signal that would be invisible to any classical image-processing approach. The models, refined iteratively across the challenge's competition rounds, learned to surface characters from noise that human annotators could not reliably parse unaided.

Luke Farritor, a 21-year-old competitor, became the first person in roughly two millennia to read a word from within an unopened scroll — an intermediate milestone that validated the ink-detection approach before the pipeline was mature enough to handle full documents. His result fed directly into the broader effort.

What This Opens

The Herculaneum library is estimated to contain hundreds of scrolls, the vast majority still unread. Most of what survived from the ancient philosophical and literary canon — particularly Epicurean philosophy, given the library's likely ownership — exists only in fragmentary form or through later copies subject to scribal drift. A scalable pipeline for reading sealed scrolls without physical intervention changes the calculus on what may still be recoverable.

The passages decoded in February 2024 were attributed to the Epicurean philosopher Philodemus, touching on the topic of pleasure — unremarkable by the standards of what we already knew of his work, but that is almost beside the point. The significance is process, not content. A pipeline that works on Philodemus can work on the scroll next to it, and the one next to that.

Worth flagging: the computational pipeline as demonstrated is not yet a push-button operation. Micro-CT scanning at the required resolution is time-intensive and the resulting datasets are enormous; the ML models have been trained on a relatively small set of validated scroll segments; and the unwrapping step still requires significant human-in-the-loop verification. Scaling from one complete scroll to hundreds is an engineering and resourcing challenge that the challenge's organizers and the broader research community have not yet solved. The first full read is a proof of capability, not a production system.

Still, the trajectory here is clear in the technical sense: each iteration of the challenge has produced better models, better unwrapping algorithms, and lower error rates. The kind of incremental, competitive, open refinement that produced this result is exactly the structure that tends to compress timelines. It took roughly two years of the Vesuvius Challenge to go from first word to first complete scroll. The next milestone may not take two more.

The Herculaneum scrolls have waited nearly 2,000 years. The tools to read them, it turns out, are arriving faster than most classicists dared expect.