How Artificial Intelligence is Unlocking the Lost Library of Herculaneum

In 79 AD, the catastrophic eruption of Mount Vesuvius buried the Roman towns of Pompeii and Herculaneum under millions of tons of volcanic ash and pumice. The disaster froze the cities in time, preserving a haunting snapshot of daily life in the Roman Empire.[1][2]

Deep beneath the mud in Herculaneum, a sprawling luxury villa—believed by historians to have belonged to Julius Caesar's father-in-law—harbored a massive private library. When farmworkers stumbled upon the ruins while digging a well in the 1750s, they discovered a cache of over 800 papyrus scrolls, representing the only intact library to survive from antiquity.[1][6]

However, the intense heat of the volcanic eruption had carbonized the scrolls, effectively turning them into brittle lumps of charcoal. For centuries, attempts to physically unroll the artifacts proved disastrous, causing the delicate papyrus to crumble into dust and destroying the ancient texts forever.[3][5]

The scrolls remained an unreadable, tantalizing mystery until the advent of modern computer vision. Brent Seales, a computer scientist at the University of Kentucky, spent years developing a non-invasive technique called "virtual unwrapping" to peer inside fragile historical documents.[6][7]

Seales's method relies on high-resolution X-ray computed tomography (CT) scans, often performed at powerful particle accelerators like the United Kingdom's Diamond Light Source. These scans create a meticulous three-dimensional map of a scroll's interior without ever touching the physical object.[2][3]

But the Herculaneum scrolls presented a unique and seemingly insurmountable hurdle. Unlike the metallic inks used in other ancient manuscripts, the Romans wrote with a carbon-based ink. Under an X-ray, the carbon ink and the carbonized papyrus possess the exact same density, rendering the text completely invisible to traditional imaging.[4]

To solve this bottleneck, Seales teamed up with Silicon Valley tech entrepreneurs Nat Friedman and Daniel Gross to launch the Vesuvius Challenge in March 2023. They released terabytes of 3D CT scan data to the public and offered over $1 million in crowdsourced prizes to anyone who could decipher the hidden text.[4][7]

The digital pipeline required to read a scroll begins with a painstaking process known as "segmentation." Researchers must meticulously trace the crinkled, warped, and compressed layers of papyrus within the 3D scan, effectively unrolling the geometry of the scroll in a digital environment.[4][6]

The breakthrough in reading the invisible ink came when a contestant noticed a faint "crackle pattern" on the surface of the segmented digital papyrus. Though invisible to X-rays, the carbon ink subtly altered the physical texture of the papyrus as it dried and burned two millennia ago, leaving a microscopic 3D topological footprint.[4]

Luke Farritor, a college student and former SpaceX intern, realized this microscopic texture could be detected by artificial intelligence. He trained a machine learning model to recognize the crackle pattern, teaching the algorithm to highlight areas where ink was present and making the invisible visible.[4][5]

In late 2023, Farritor's model successfully revealed the first legible word from an unopened Herculaneum scroll: "porphyras," the ancient Greek word for purple. This milestone proved that the invisible ink could finally be read, sending shockwaves through the fields of computer science and classical history.[1][5]

The competition rapidly accelerated following the initial discovery. Farritor teamed up with two other students, Youssef Nader and Julian Schilliger, to refine the AI and segmentation tools. By early 2024, the trio had deciphered 15 columns of text—over 2,000 characters—claiming the $700,000 grand prize.[1][5]

Papyrologists translated the newly revealed text, attributing it to the Epicurean philosopher Philodemus. The author discusses the nature of pleasure, debating whether the scarcity or abundance of goods, such as food and music, affects the enjoyment derived from them.[2][7]

The technology has only improved since the grand prize was awarded. In early 2025, researchers applied the refined AI pipeline to a different scroll, known as PHerc. 172, which is currently housed at Oxford's Bodleian Libraries.[2][3]

This scan yielded more recoverable text than any previously analyzed Herculaneum papyrus. Scholars identified multiple columns of clean text, including the Greek word "diatrope," meaning disgust, which appeared several times across the cleanly separated layers.[2][3]

Despite these monumental successes, uncertainties remain. The AI models are highly effective on well-preserved sections, but heavily compressed or torn interior layers still pose a significant challenge. If the algorithms are not strictly constrained by the underlying 3D papyrus fibers, they risk hallucinating letter shapes.[4][6]

The stakes for historical preservation are immense. Historians estimate that only 3 to 5 percent of ancient Greek texts have survived to the modern era. Because the Herculaneum library is the only intact library known from the classical world, experts believe thousands more scrolls may still be buried in unexcavated portions of the villa.[1][7]

What began as a tragic natural disaster 2,000 years ago has been transformed into a triumph of modern computer vision. The Vesuvius Challenge has not only resurrected a lost library but provided a digital blueprint for recovering fragile, forgotten voices across antiquity.[2][7]