We’re finally reading the secrets of Herculaneum’s lost library

In a particle accelerator deep underground, theoretical physicist Giorgio Angelotti is diligently performing his tasks. He secures a black cylinder on its mount, affixes it tightly, and methodically conducts safety checks before retreating from the chamber known as “the hatch.” “It is crucial to ensure that no one is in the hatch before closing the door,” he emphasizes. “We have to prevent any risk of death.”

He is on the verge of bombarding a sample with an intense beam of X-rays. You might imagine the target being a sophisticated new material or a delicate crystal, but this situation is quite different. The object encapsulated within the cylinder is far from pristine; it resembles a distorted piece of ancient charcoal.

In fact, it is an invaluable artifact: a 2,000-year-old papyrus scroll, burned beyond recognition in the devastating eruption of Vesuvius in AD 79. This fragment is part of the Herculaneum papyri—an extensive collection of scrolls that are too fragile for manual opening, leaving their contents largely unknown. By utilizing particle accelerators, artificial intelligence, and a skilled team of coders dispersed across the internet, Angelotti and his group are beginning to make these carbonized relics communicate. They may soon unveil entire lost works of Greek philosophy or writings from the earliest Christians.

Unearthed near Angelotti’s hometown of Naples, Italy, in the 1750s, these scrolls originate from a partially excavated 1st-century BC villa located in Herculaneum. The town, a lesser-known neighbor of Pompeii, was once a favored seaside retreat for wealthy Romans. The opulent villa is believed to have belonged to the Roman senator Lucius Calpurnius Piso Caesoninus—Julius Caesar’s father-in-law.

Some of the original 900 scrolls are attributed to the philosopher Philodemus of Gadara, known for disseminating Epicurean philosophy from Greece to Italy. Classicist David Blank from the University of California, Los Angeles, indicates that Philodemus was part of Piso’s cercle, a group whose intellectual caliber highlighted the senator’s status. Consequently, Piso became a benefactor of Philodemus’s works, ensuring that many of his philosophical texts, including rare early drafts, became part of Piso’s collection.

The Herculaneum papyri

Though Piso and Philodemus had been deceased for decades at the time of Mount Vesuvius’s eruption, the library endured. As hot ash and mud consumed Herculaneum, the heat dried out the scrolls, carbonizing them rather than incinerating them. “Their carbonized state is the only reason we still possess them,” explains papyrologist Federica Nicolardi from the University of Naples Federico II. Typically, papyrus survives solely in arid environments; other European examples deteriorated centuries ago.

Unfortunately, the Piso collection has diminished significantly over time. The layers of papyrus are tightly fused, and early initiatives to unravel them resulted in many being crushed, sliced, and treated in ways that would make any papyrologist cringe. Starting in the 1750s, the scrolls’ initial custodian, Camillo Paderni, forcibly extracted their interiors, leaving only the outer layers. “He would take the roll, cut through it… then use the blunt end of a knife to pound the center of the roll into dust,” recounts Blank.

Later, Antonio Piaggio, a manuscript restorer from the Vatican Library, applied his own invention to some of the scrolls. By attaching each scroll to a sheet made of animal calligraphy using fish-based glue, he managed to carefully unroll about 18 of them. These initial attempts produced several volumes of comprehensible texts, revealing at least some works by Philodemus. Yet the majority of these charcoal lumps remained unread in Naples’ National Library.

This state persisted until Brent Seales from the University of Kentucky emerged on the scene. Having experienced the first wave of digitization, as the internet transformed into a trove of knowledge, he became captivated by the notion that vital pieces of this library might be overlooked due to physical damage. “The idea that technology could reproduce or extract novel information from damaged items captivated me,” he recounts.

In 2000, Seales utilized 3D scanning alongside computer software to digitally uncrumple and flatten pages from fire-damaged medieval texts collected by Sir Robert Cotton, now part of the British Library’s foundational collection. Some volumes, however, remained too unyielding to be opened, making them unfit for restoration via conventional imaging techniques reliant on visible light. This prompted Seales to consider whether methods used to examine internal structures of living beings could similarly be applied to books.

Upon directing X-rays at a book within the Cotton collection, he began to see the ink emerge in a manner akin to how bones appear on black-and-white imaging. Immediately, he sought access to other collections housing unopened texts, and the most renowned example came to mind: the Dead Sea Scrolls. However, when Seales shared his idea with conservators, they firmly rejected it. His focus then shifted to the Herculaneum scrolls, spurred by insights from classicist Richard Janko from the University of Michigan, who had explored the contents of some previously opened scrolls.

These specific papyri presented unique challenges. Unlike medieval book authors who utilized metallic inks, Philodemus and his contemporaries frequently penned texts using soot-based ink, complicating efforts to distinguish the carbon-based ink from the surrounding carbonaceous scrolls. Initial trials with a small CT scanner in 2009 did not yield any visible ink.

Many Hebrew and Egyptian scribes utilized metallic inks that were easier to image. By 2015, Seales successfully deciphered hidden text inside a charred 4th-century AD Hebrew scroll. Soon after, a European team, including Verena Lepper from Berlin’s Egyptian Museum and Papyrus Collection, utilized X-ray scans to reveal the words “oh Lord” within an ancient papyrus package from the Nile island of Elephantine. However, scans of the Herculaneum scrolls had yet to yield a single word.

Digitizing the unwrapping process was equally challenging. The tightly packed papyrus layers made virtual separation difficult. Nicolardi elaborates, “If the software cannot differentiate between layers, you produce results akin to those from mechanically opened scrolls,” leading to mangled text.

By this time, AI technology was advancing, enabling machines to recognize features undetectable to humans. It became evident that scans of the Herculaneum papyri were indeed capturing ink, but only a suitably configured AI could see it. In 2019, Seales and his colleagues successfully demonstrated this on unrolled fragments and replicas inscribed with carbon ink, which allowed them access to the Diamond Light Source in Oxford, UK. They employed the accelerator as an enhanced CT scanner, capturing internal images of rolled-up intact papyri. Yet the scrolls continued to elude them. Seales’s student Stephen Parsons developed AI software to identify ink in these high-resolution scans, but it struggled to detect anything beyond mere traces.

At that point, a significant shift occurred. Seales connected with tech investor Nat Friedman, the former CEO of GitHub, intending to secure further research funding. However, Friedman pitched a different idea: launch a public challenge to see if someone could write a program capable of reading the scrolls. Seales was initially hesitant. This cash-for-code challenge was standard in the tech sector but an alien concept in academia, requiring him to make scan data and Parsons’s algorithms available to a broader audience. “We realized our reluctance stemmed from concerns about recognition, which wasn’t a strong reason,” Seales acknowledged.

The Vesuvius Challenge

Thus, in March 2023, the Vesuvius Challenge emerged. Any resultant solutions would be made public, and the code would be shared for the team and others to build upon, speeding up progress. The initiative proved successful; by Christmas, the challenge’s Discord channel boasted over 1,000 users.

Angelotti was among the new participants. Fresh from earning a doctorate in AI, he had little prior knowledge of the Herculaneum scrolls, despite being a native of Naples. However, as he delved deeper, they captivated him. Balancing consultancy and launching an AI start-up, he immersed himself in digitized papyrus sheets online. Despite lacking papyrology expertise, the effort paid off, yielding cash prizes totaling $20,000 for accelerating image processing and ultimately a job offer. Now the research project lead for the Vesuvius Challenge, Angelotti describes reading the scrolls as “a quest to restore the cultural heritage of my homeland.”

Meanwhile, students began gaining prominence. In December 2023, ink-detection algorithms crafted by Youssef Nader and Luke Farritor uncovered approximately 2,000 Greek characters. Nader trained the AI to recognize ink by focusing on fragments of scrolls with exposed papyrus surfaces. Concurrently, Farritor identified the first word, porphyras (purple), from an unopened scroll through a distinct AI model trained on sections showing faint, visible “crackle” patterns associated with inked regions.

Their collaboration with Julian Schilliger, a student at ETH Zürich in Switzerland, who had successfully stitched together digital papyrus sheets made from pixels, led to improved outcomes and recognition in a peer-reviewed papyrology paper. The translated text revealed ancient reflections on food, music, and pleasure, where the author mused on the essence of a meaningful life.

Their contributions earned them the grand prize of $700,000 in the Vesuvius Challenge during 2023—as well as a cake resembling Mount Vesuvius (complete with a scroll), baked by Nader’s family in Egypt. Nader later joined the challenge team to continue researching ink detection. Although progress remains ongoing, there’s still variation in the inks used across different scrolls. Ultimately, the team aims to develop a fast, general ink-detection software capable of functioning across all types, “so we can upload a scroll scan and obtain the text,” Nader stated.

The challenge of unrolling is yet to be fully resolved. Initially, the inked surfaces of the papyrus layers required meticulous mapping to flattened sections, often handled by humans. However, with the support of community members like Schilliger, the team now increasingly relies on AI to undertake these tasks, which should provide swifter results.

Will breakthroughs for these scrolls benefit other ancient papyri as well? “I don’t believe there is a single solution, nor is one necessary,” says Lepper, whose work on the Elephantine papyri utilized traditional, non-AI software. Each collection exhibits its own characteristics, she notes. For instance, while Elephantine papyri aren’t charred, many are folded rather than rolled, complicating the process.

Unveiling hidden writing in historical manuscripts remains a formidable challenge. Yet, for the Vesuvius Challenge, progress continues to unfold “as a direct response to the competition,” remarks Seales, his initial uncertainties now fading. Both Seales and Angelotti are hopeful that a time will come when extracting text is as simple as clicking a button and letting the software handle the rest. However, for now, numerous scrolls await scanning, translating to more hours spent in the controlled environments of particle accelerators.

When New Scientist spoke with Angelotti in mid-July, he detailed scanning over 30 Herculaneum scrolls at the Diamond Light Source and the European Synchrotron Radiation Facility with “the hatch.” He was also conducting crucial experimental tasks, and early results suggested that scanning at a higher resolution might allow AI to detect common ink features across all scrolls. If this proves effective, the whole collection could become readily accessible, though it would involve scans taking roughly six times longer than usual—resulting in extended hours in the control room.

Meanwhile, the Vesuvius Challenge team has been gearing up to share more data with its community of coders, and achievements continue to mount. In May 2025, computer science graduates Marcel Roth and Micha Nowak at the University of Würzburg in Germany modified medical imaging software to successfully read the first-ever title from within the scrolls, which won them $60,000. Roth remarked that the pair became so engrossed in the competition that they temporarily suspended their university studies for nearly three months.

And what was the title? Philodemus, On Vices. “We were thrilled to confirm it was indeed Philodemus,” noted Angelotti, affirming the AI wasn’t hallucinating. It’s unlikely that report will be the last we hear of Philodemus, as it seems that most of the explored scrolls originate from the philosophical section of Piso’s grand library.

In the Bay of Naples, there are potentially many more scrolls yet to be discovered. A significant portion of the villa remains unexcavated, buried beneath 20 meters of volcanic debris and impeded by local bureaucratic issues. The New Testament places Paul the Apostle in the area around AD 50, prior to his execution approximately 15 years later. Could these scrolls chronicle his movements before Vesuvius’s catastrophic eruption? Perhaps, “if Herculaneum’s library had a current events section,” Seales joked. Until recently, however, delving for such long-lost treasures seemed futile, as unlocking their secrets was an impossibility. With these barriers now dissolving, it presents a compelling case for resuming archaeological efforts.