AI Decodes Ancient Herculaneum Scrolls Without Unrolling

For nearly 2,000 years, the library of the Villa of the Papyri lay buried under 20 meters of volcanic mud. When Mount Vesuvius erupted in 79 AD, the intense heat carbonized hundreds of papyrus scrolls, turning them into fragile lumps of charcoal. Attempting to open them physically has always resulted in their destruction. Now, a combination of particle physics and artificial intelligence has achieved the impossible. Without touching the papyrus, researchers have read entire paragraphs from inside these closed scrolls.

The Vesuvius Challenge

The breakthrough is the result of a global competition known as the Vesuvius Challenge. Launched in March 2023, this project was spearheaded by Nat Friedman, the former CEO of GitHub, alongside Daniel Gross and computer scientist Brent Seales from the University of Kentucky. They released high-resolution CT scans of the scrolls and offered over $1 million in prizes to anyone who could decipher the text.

The primary obstacle was that the ink used by the Romans was carbon-based (charcoal and gum), and the scrolls themselves were turned into carbon by the volcano. Unlike iron-gall ink used in medieval manuscripts, Roman ink does not show up clearly on standard X-rays because it has the same density as the paper it is written on.

To solve this, Brent Seales developed a technique called “virtual unwrapping.” However, the ink remained difficult to see. The Vesuvius Challenge crowdsourced the solution, inviting machine learning experts to train AI models to detect subtle textural changes on the surface of the papyrus where the ink was applied.

How Virtual Unwrapping Works

The process of reading these scrolls involves three distinct, high-tech steps. It is a mix of heavy industrial hardware and advanced software code.

• Scanning: The scrolls were scanned at the Diamond Light Source in the United Kingdom. This particle accelerator produces X-rays 10 billion times brighter than the sun. This allows for scans with a resolution of 4 to 8 micrometers, capturing the tiniest details of the internal structure.
• Segmentation: Before reading, the computer must understand the geometry of the scroll. The scrolls are not rolled perfectly; they are crushed, twisted, and deformed. Programmers use software to trace the individual layers of papyrus through the 3D data block. This flattens the rolled surface into a 2D sheet digitally.
• Ink Detection: This is where the AI takes over. The machine learning models look for a “crackle pattern.” When the ink dried on the papyrus 2,000 years ago, it created a microscopic disruption in the fibers. The AI analyzes the texture of the flattened 2D images and predicts, pixel by pixel, where ink is present.

The Grand Prize Winners and the First Word

In October 2023, a 21-year-old computer science student named Luke Farritor became the first person in two millennia to see a complete word from inside an unopened scroll. He trained a model that identified the Greek word porphyras, which means “purple.” This discovery proved the concept worked and won him the “First Letters” prize of $40,000.

Following this initial success, the competition heated up. In February 2024, the Vesuvius Challenge awarded the $700,000 Grand Prize to a team of three super-coders:

1. Youssef Nader: An Egyptian biorobotics student in Berlin.
2. Luke Farritor: The student from Nebraska who found the first word.
3. Julian Schilliger: A robotics student from Zurich who built tools to automate the segmentation process.

Together, their combined architectures revealed over 2,000 characters from four separate columns of text. This constituted about 5% of the scroll, meeting the criteria for the grand prize.

What Do the Scrolls Say?

The text revealed by the AI is not just a receipt or a laundry list. It is a philosophical work. Papryologists and scholars reviewing the output believe the text was likely written by the Epicurean philosopher Philodemus, who was the teacher of the Roman poet Virgil.

The decoded passages discuss the nature of pleasure. The author argues against the idea that scarcity increases value. In a rough translation, the text asserts that goods which are abundant can be just as pleasant as those that are rare. The text specifically mentions food, music, and the color purple.

One translated section reads: “As too in the case of food, we do not right away believe that things that are scarce are absolutely more pleasant than those which are abundant.” The author ends the passage by taking a jab at unnamed ideological adversaries, promising to have “nothing to say” about them.

The Future: Scanning the Rest of the Library

The success of the 2023 Vesuvius Challenge has unlocked a new era for archaeology. There are approximately 800 scrolls from the Herculaneum library stored in Naples today. However, historians believe the Villa of the Papyri has two more levels that have never been excavated. It is possible that thousands of scrolls remain underground.

The implications are staggering. Most of classical literature has been lost. We have only a fraction of the works of Sophocles, Aeschylus, and Livy. If this technology can be scaled, we might recover:

• Lost plays and poems from the Greek Golden Age.
• Early Roman history written by contemporary witnesses.
• Documents detailing the political life of the Roman Empire in the 1st Century.

The goal for the next phase of the project is to scale the process. The organizers have set a new challenge to read 90% of a scroll, rather than just 5%. This will require automating the segmentation process, which is currently the most labor-intensive part of the workflow.

Frequently Asked Questions

Why can’t we just unroll the scrolls carefully? The scrolls have been carbonized by the heat of the pyroclastic flow. They are essentially lumps of charcoal. Any attempt to physically pull them apart causes them to crumble into dust. Early attempts in the 18th and 19th centuries destroyed many scrolls, yielding only fragments of text.

What language are the scrolls written in? The scrolls deciphered so far are written in Ancient Greek. However, standard Roman libraries typically had two sections: one for Greek and one for Latin. Scholars are hopeful that future scans will reveal Latin texts, which could contain lost works of Roman history and literature.

How accurate is the AI translation? The AI does not translate; it reveals the visual imagery of the letters. It produces an image that looks like a black-and-white scan of handwriting. Papyrologists (experts in ancient manuscripts) then read that imagery just as they would a physical paper. The clarity of the AI-generated images has been praised by scholars as being surprisingly readable.

Who owns the physical scrolls? The scrolls are owned by the Italian government and are housed at the National Library of Naples. The Vesuvius Challenge team works in partnership with the library and the Institut de France to access the scrolls for scanning.