How AI is Reading the 2,000-Year-Old Carbonized Scrolls of Herculaneum

In 79 AD, the eruption of Mount Vesuvius unleashed a torrent of superheated gas and ash that obliterated the Roman cities of Pompeii and Herculaneum. Amid the destruction, a massive luxury estate—believed to belong to Julius Caesar's father-in-law—was buried under twenty meters of volcanic mud. Inside this estate, known today as the Villa of the Papyri, lay a vast library of ancient texts. The intense heat of the pyroclastic flow flash-fried the library, instantly carbonizing the papyrus scrolls. For nearly two millennia, the texts sat in the dark, perfectly preserved but transformed into brittle lumps of charcoal.[1][4][5]

When excavators finally unearthed the villa in the 1750s, they discovered roughly 1,800 of these blackened cylinders. It was the only intact library to survive from the classical world. However, the discovery quickly turned into a tragedy of preservation. The carbonized papyrus was so fragile that early attempts to physically unroll the scrolls caused them to shatter into dust. While a few were painstakingly peeled apart over decades, destroying much of the text in the process, more than 600 scrolls were deemed completely unopenable and locked away as indecipherable curiosities.[1][4][5]

For centuries, the Herculaneum papyri represented one of archaeology's most tantalizing locked doors. Scholars knew the scrolls contained lost works of Greek and Roman philosophy, but accessing them seemed physically impossible. The paradigm only began to shift in the early 21st century, driven by advancements in medical and industrial imaging. Dr. Brent Seales, a computer scientist at the University of Kentucky, pioneered a technique called "virtual unwrapping," which uses X-ray microtomography (micro-CT) to create 3D volumetric scans of closed documents.[2][4][5][7]

In 2015, Seales and his team achieved a landmark victory by virtually unwrapping the En-Gedi scroll, a charred Hebrew parchment from Israel, revealing text from the Book of Leviticus without ever opening the physical artifact. But when the team applied the same X-ray techniques to the Herculaneum scrolls, they hit a wall. The En-Gedi ink contained metal, which glowed brightly on an X-ray. The Roman writers at Herculaneum, however, used a carbon-based ink made of soot and water. On an X-ray, the carbon ink was completely indistinguishable from the carbonized papyrus it was written on.[1][2][5]

To solve the ink problem, researchers realized they needed unprecedented imaging resolution. They took the fragile scrolls to particle accelerators, such as the Diamond Light Source in the United Kingdom, which can generate X-ray beams billions of times brighter than the sun. At these microscopic resolutions, a new detail emerged: the carbon ink didn't change the chemical density on the scan, but it did slightly alter the physical geometry of the papyrus. The ink coated the ancient fibers, creating a faint 3D texture—a "crackle pattern"—that was invisible to the human eye but theoretically detectable by software.[1][3][5][6]

Recognizing that the bottleneck had shifted from archaeology to computer science, Seales teamed up with Silicon Valley entrepreneurs Nat Friedman and Daniel Gross. In March 2023, they launched the Vesuvius Challenge, releasing terabytes of high-resolution 3D scroll scans and open-sourcing the initial machine-learning algorithms. They offered a $1 million prize pool to anyone in the world who could improve the AI models and extract readable text. The crowdsourced approach transformed a niche academic pursuit into a global technology race.[1][2][4][5][7]

The breakthrough came astonishingly fast. In late 2023, Luke Farritor, a 21-year-old college student and SpaceX intern, trained a machine-learning model specifically on the faint crackle patterns discussed in the project's online forums. His software successfully detected ink traces deep inside a closed scroll, revealing the Greek word "porphyras," meaning purple. It was the first time in 2,000 years that a word had been read from an unopened Herculaneum scroll.[1][4]

The momentum accelerated into 2024. Farritor teamed up with Youssef Nader, an Egyptian graduate student in Germany, and Julian Schilliger, a Swiss robotics engineer. Together, they won the $700,000 Grand Prize by developing a sophisticated pipeline that virtually unrolled the 3D scans and used AI to identify over 2,000 Greek characters across 15 columns of text. The decoded passages were revealed to be a previously unknown work by Philodemus, an Epicurean philosopher, discussing the nature of pleasure, music, and food.[1][3][4]

The success of the Vesuvius Challenge proved that the entire Herculaneum library was theoretically readable. In 2025, the project expanded its scope, scanning additional scrolls housed at the Bodleian Libraries in Oxford. Using the refined AI pipeline, researchers quickly generated images of the inside of scroll PHerc. 172, identifying words like "diatrope," meaning disgust. Interestingly, the Bodleian scroll exhibited a slightly different ink chemistry, suggesting that different scribes or time periods might require specialized AI models.[3][5][6]

The current frontier of the project is automation. While the 2024 Grand Prize proved the concept, the process of digitally segmenting the tangled, crushed layers of papyrus remains highly manual and labor-intensive. Computer scientists are now competing to develop algorithms that can automatically trace the continuous surface of a scroll through the chaotic 3D data, a process known as "auto-segmentation." Solving this will allow researchers to move from reading isolated patches of text to deciphering entire scrolls in a matter of days.[1][7]

The implications for classical history are staggering. Currently, the surviving corpus of ancient Greek and Roman literature represents only a tiny fraction of what was written; the rest was lost to decay, war, and religious suppression. The Villa of the Papyri is the only place on Earth where a comprehensive ancient library remains intact. If the automated AI pipeline can decode the remaining 600 scrolls, it could fundamentally alter our understanding of the ancient world.[1][4][7]

Furthermore, archaeologists believe that the 1,800 scrolls recovered in the 18th century represent only a portion of the villa's collection. The main library of the estate may still lie buried under the unexcavated sections of Herculaneum. With the knowledge that any recovered scroll can now be read safely, there is renewed momentum to resume excavations at the site, potentially uncovering thousands of additional texts.[5][7]

The Vesuvius Challenge stands as a landmark example of interdisciplinary collaboration. It required the expertise of classicists to translate the Greek, physicists to operate the particle accelerators, and a global community of computer scientists to build the machine-learning models. By treating a 2,000-year-old archaeological mystery as an open-source software problem, the project bypassed decades of traditional academic siloing.[2][6][7]

As the AI models continue to improve, scholars are eagerly awaiting the next batch of translations. While the texts decoded so far have been Epicurean philosophy, the library could hold lost plays by Sophocles, missing histories of the Roman Republic, or early scientific treatises. The ashes of Mount Vesuvius, which brought death and destruction to the Bay of Naples, inadvertently created a perfect time capsule. Now, artificial intelligence is finally providing the key to open it.[1][4][7]