The Library Made of Charcoal: How X-Ray Light and AI Are Reading the Burned Scrolls of Herculaneum

The Hook: A Book You Can Only Read by Destroying It

Imagine you own the only surviving book from a lost library. It has endured for two thousand years — but only because a volcano turned it into a lump of charcoal. The papyrus has not rotted, molded, or crumbled away; it has been carbonized, petrified into a black, brittle roll, as fragile as a burnt piece of toast. If you try to unroll it to read what it says, it disintegrates into splinters in your hands. The knowledge is there. It is physically present, letter by letter. And it is unreachable, because every attempt to reach it destroys it.

That is precisely the situation of the Herculaneum papyri: roughly 1,800 carbonized scrolls buried in 79 AD by the eruption of Mount Vesuvius under a stream of glowing ash — the only library from Greco-Roman antiquity that has survived intact as a whole to this day. For two and a half centuries their contents were held captive. Anyone who wanted to open them risked losing them forever, and hundreds were indeed destroyed in early attempts to unroll them.

And then, between 2023 and 2026, something happened that had long been considered impossible. An international team of physicists, computer scientists, students, and classicists began to read these scrolls without opening them. They shone the world's brightest X-ray light through the sealed lumps, reconstructed on a computer the tightly wound layer of papyrus inside, "unrolled" it virtually into a flat surface, and then trained neural networks to reveal an ink that is practically invisible — because it consists of the same carbon as the carbonized papyrus on which it sits. In October 2023 the first word fell. In June 2026 the first complete scroll was read from beginning to end.

Why should this interest you, beyond the fascination with sunken libraries? Because three threads converge here that shape your professional life: non-invasive reconstruction of a state you are not allowed to observe directly; machine learning on a tiny, noisy signal that a human being can no longer detect with the naked eye; and an open, prize-driven competition model that cracked a two-thousand-year-old problem in less than three years, after decades of classical scholarship had failed at it. This is not antiquarianism. It is a blueprint for how to pull a signal out of noise when the truth is present but hidden.

Part 1: The Villa of the Papyri — Antiquity's Only Library

An Ash Flow as a Preservative

On August 24, 79 AD — the date is disputed in scholarship, and there is some evidence for an autumn timing — Vesuvius erupted and destroyed not only Pompeii but also the smaller, wealthier coastal town of Herculaneum. While Pompeii suffocated under pumice and ash, Herculaneum was struck by something else: pyroclastic flows, avalanches of superheated gas and rock at temperatures of several hundred degrees Celsius. This heat killed instantly, but it did something paradoxical to organic material. Wood, food, and papyrus were not burned in the sense of being reduced to ash but, in the absence of oxygen, carbonized — converted into carbon, the way wood becomes charcoal in a kiln. What would normally have rotted away within decades was transformed into a chemically stable, if extremely brittle, form and then sealed under as much as twenty meters of solidified volcanic material.

This is the cruel irony that runs through the whole story: the very catastrophe that destroyed the library is also the only reason it was preserved from antiquity. No other significant collection of book rolls from the classical world has physically survived. Everything else we possess of ancient literature we know only because it was copied by hand, again and again, over the centuries. Herculaneum is the only library whose books themselves are still there.

The Discovery and the Resident

In the 1750s, workers digging tunnels for antiquities on behalf of the Bourbon king Charles of Naples came upon a luxurious villa at the edge of the town. It contained marble and bronze statues of such quality that the building soon became known as the Villa dei Papiri — Villa of the Papyri — named after its true treasure: the scrolls. At first the excavators mistook the carbonized cylinders for firewood or lumps of coal and threw some away, until it was realized that they were books. In the end, some 1,800 rolls were recovered — the private book collection of an educated Roman.

The contents, as far as they were known, point strongly to one particular mind: Philodemus of Gadara (c. 110–35 BC), a Greek philosopher of the Epicurean school who lived and worked in Italy. A large portion of the texts read so far comes from him or his circle — treatises on ethics, on music, on rhetoric, on the gods, on death, on the art of living well. The villa probably belonged to a Roman aristocrat, possibly Lucius Calpurnius Piso Caesoninus, the father-in-law of Julius Caesar, in whose circle Philodemus moved as a kind of resident philosopher. What lies under the ash, then, is not just any library but a window into the intellectual life of the late Roman Republic — and potentially into lost works whose existence we know only from passing mentions.

The Destructive Attempts to Open Them

The temptation to read these books was irresistible, and the early methods were brutal. Some rolls were simply broken open or split lengthwise. The great advance of the eighteenth century was the machine of Antonio Piaggio, a conservator from the Vatican: a contraption of threads and weights that pulled a roll apart millimeter by millimeter over weeks and months. This occasionally succeeded — legible fragments were indeed recovered — but the price was high. Many rolls fell apart in the process, and what remained was often a mosaic of fragments whose original order had been lost.

It was precisely the most tightly wound, best-preserved rolls that most stubbornly resisted every mechanical opening. For over two hundred years, therefore, a hard core of hundreds of rolls remained unopened — too valuable to risk, too brittle to touch. They waited for a technique that did not yet exist: a way to look inside without touching the outside.

Part 2: Why Carbon on Carbon Is Invisible — the Core Technical Problem

The decisive obstacle is a physical one, and it is trickier than it first sounds. The obvious idea is: if you cannot open the roll, you simply shine X-rays through it, the way a doctor X-rays a bone. A computed-tomography (CT) scanner produces a three-dimensional density image — and it can indeed make the wound layer of papyrus inside visible, coil by coil.

The problem is not the papyrus. The problem is the ink.

Ancient scribes mostly used a carbon ink — essentially soot, fine carbon, mixed with water and a binder such as gum arabic. On fresh, light-colored papyrus this produces a rich black on beige, perfectly legible. But once Vesuvius had turned the papyrus itself into carbon, you now have carbon sitting on carbon. To an X-ray beam, which distinguishes matter by its density, ink and substrate are nearly identical. The letters have no density contrast against their background — they are simply invisible in the CT scan, the way white writing on white paper vanishes to the eye.

A contrast makes the scale of this problem clear. In 2015 the team around Brent Seales at the University of Kentucky achieved a spectacular breakthrough with the En-Gedi scroll, a carbonized Hebrew parchment from the shore of the Dead Sea. It turned out to be the beginning of the biblical book of Leviticus — one of the oldest known Hebrew biblical texts. Why did this go so "easily"? Because the scribes of En-Gedi used a metal-based ink, probably containing iron. Metal is far denser than parchment in an X-ray image; the letters practically lit up. With En-Gedi, the writing merely had to be geometrically disentangled, not first made visible.

Herculaneum lacks exactly this luxury. Its carbon ink provides no glow. I am of the opinion that this single circumstance — the absence of a density difference between ink and substrate — explains why Herculaneum was the far harder nut, and why the solution could not come from physics alone but required machine learning. One had to teach an algorithm to recognize, in the X-ray data, the tiny, indirect traces that the ink leaves behind — not its density, but subtle differences in texture, in surface structure, in the pattern of the papyrus where soot once lay. A signal so faint that a human does not notice it when looking directly.

Part 3: Virtual Unwrapping — Brent Seales's Twenty-Year Vision

An Idea Ahead of Its Time

The conceptual foundation was laid, long before the great success, by a single researcher: W. Brent Seales, a professor of computer science at the University of Kentucky. Since the early 2000s he had pursued an idea he called virtual unwrapping: one need not physically open a damaged document to read it — one can capture its interior with imaging techniques and reconstruct it purely computationally. "Virtual unwrapping was born from a vision I had where we might be able to explore the interior of something without having to physically open it," he later described it.

The method breaks down into three steps that still form the basic framework today. First, segmentation: in the three-dimensional X-ray scan, one must find the tightly wound surface of papyrus and trace it coil by coil — following a single, meters-long spiral through a compact volume in which the layers almost touch. Second, unwrapping and flattening: the curved surface thus identified is mathematically transformed into a flat, readable page without distorting the geometry. Third, texturing, or ink detection: onto the flat surface one projects the signal that betrays the ink, thereby making the letters visible.

The Proof: En-Gedi

With the En-Gedi scroll, Seales proved in 2015 that the chain works — at least when the ink is metallic and therefore visible in the X-ray image. From a carbonized lump that no human could ever have opened, his method retrieved coherent, readable biblical text. This was the proof of feasibility that made everything that followed conceivable at all. But En-Gedi, measured against Herculaneum, was the easy case.

The Leap to the Synchrotron

For Herculaneum, Seales needed sharper eyes. An ordinary laboratory CT does not have the resolution to cleanly separate the wafer-thin, densely packed layers of a Herculaneum roll, let alone to catch the fine textural traces of carbon ink. So in 2019 he went to a synchrotron facility: the Diamond Light Source near Oxford, a ring-shaped particle accelerator that generates extremely intense, precisely controllable X-ray light. There he scanned two complete rolls from the holdings of the Institut de France in Paris using X-ray phase-contrast tomography at a resolution unattainable in a home laboratory. A later research consortium, the EduceLab, founded in 2021 with 14 million dollars from the U.S. National Science Foundation, gave this work a permanent home.

For the first time, high-resolution volumetric data of the roll's interior was now available. Yet the ink remained invisible to the naked eye, and the volumes of data were enormous. Seales faced a classic scaling problem: the decisive task — learning to pull the ink out of the noise — was too big for a single laboratory. It was at exactly this point that an unusual decision accelerated everything.

Part 4: The Vesuvius Challenge — a Prize Opens the Problem to the World

A Laboratory Becomes a Movement

In March 2023, Seales did something rare in the academic world: he opened his data and his software to everyone. Together with the tech entrepreneur Nat Friedman (the former CEO of GitHub) and the investor Daniel Gross, he launched the Vesuvius Challenge — an open, donation-funded competition with large cash prizes. The idea: instead of solving the problem behind laboratory doors, you put the X-ray scans publicly online, offer a series of clearly defined prizes, and let the whole world of tinkerers, students, and machine-learning enthusiasts work on it. The approach recalls the great historical prize competitions with which, for example, the longitude problem of seafaring was solved: you do not buy working time, you buy a result.

The gamble paid off spectacularly. Around a Discord server, an international community formed within weeks, sharing tools, intermediate results, and training data — an open, almost swarm-like research process.

The First Word: Purple

The first great breakthrough came in October 2023 from Luke Farritor, a computer science student from Nebraska who worked on the Challenge on the side. He trained a neural network to recognize the textural traces of the ink in the scans, and with it made a first coherent word visible: ΠΟΡΦΥΡΑϹ — porphyras, the Greek word for purple. Shortly afterward, Youssef Nader, an Egyptian doing his doctorate in Germany, independently confirmed and extended the result. For the first time in nearly two thousand years, a human had read a word from an unopened Herculaneum roll. It was a single word — but it proved that the ink was there in the signal and that a learning system could pull it out.

The Grand Prize: the First Passage

Only a few months later, in February 2024, the Grand Prize of 700,000 US dollars was awarded. A team of Youssef Nader, Luke Farritor, and Julian Schilliger had made roughly 2,000 Greek letters across about 15 columns of a scroll (known to the community as "Scroll 1," officially PHerc. Paris 4) legible — coherent, interpretable text. In content it is an Epicurean treatise on pleasure and the senses: the author — in all likelihood Philodemus again — discusses what makes for enjoyment of life, taking up examples such as music and food. It is the sort of text no one had ever read before, because it was contained in exactly that roll and survives nowhere else. A dead black cylinder had become a readable chapter of a book.

Part 5: From Words to Titles to Whole Scrolls (2025–2026)

The Title: On Vices

One deep practical problem remained: to place a scroll scholarly, you first want to know its title — but in antiquity that often stood at the end, in the innermost, most inaccessible part of the roll. In 2025 this hurdle fell. In the scroll PHerc. 172, held in the Bodleian Library of the University of Oxford, Marcel Roth and Micha Nowak — and, independently, the Challenge researcher Sean Johnson — deciphered the title. The work is "On Vices" (Peri kakion) by Philodemus, an ethical treatise on the failings of human beings and the virtues opposed to them. It was the first time ever that the title of a still-rolled, never-opened Herculaneum scroll had been recovered non-invasively; the discovery earned the "First Title" prize, endowed with 60,000 dollars.

The Breakthrough: an Entire Scroll, from Beginning to End

The greatest leap so far dates from June 25, 2026. The Vesuvius Challenge research team announced that it had, for the first time, completely virtually unwrapped and read the scroll PHerc. 1667 (known to the community as "Scroll 4") from beginning to end — the first Herculaneum roll available as continuous, coherent text rather than in isolated words or patches. The early, mechanical attempts to open it in the nineteenth century and again in 1969 and the 1980s had destroyed the outer layers and left only the compact inner core, about eight centimeters of an original height of some 19 to 24 centimeters. From this surviving remnant the team read the lower parts of some 22 columns — transcribed and reviewed by papyrologists.

And the content holds a surprise. It is not by Philodemus but is a Stoic philosophical treatise on ethics — on human nature, on impulse, and on the moral progress of human beings. Its final preserved column names Aristocreon, the nephew and disciple of the great Stoic Chrysippus, which, together with the language and themes, places the text in a Stoic context and dates it to the 2nd century BC. This broadens the picture of the library: it contains not only Epicurean but also Stoic works. For the first time in two thousand years, sentences like this can be read clearly again: "…we will inquire into something, but we will not grasp it, if in some way we depart from ourselves and from our own nature…"

In the same announcement the team reported two further results that secure and extend the method. In PHerc. Paris 4 (Scroll 1), an even higher resolution made the ink, for the first time, directly visible in the three-dimensional X-ray data — one can literally delineate it within the volume. Projected back onto the unwrapped page, this ink matches the text the team read for the Grand Prize in 2024 one-to-one: an independent confirmation, from better data, that the reading is real and not an artifact. And in a third scroll, PHerc. 139, the title and author attribution were recovered: the work is Philodemus, "On Gods," Book 8.

How It Works — the Pipeline at a Glance

It is worth looking at the present-day procedure as a whole, for it is a model of how physics, geometry, and machine learning mesh together. The current scans were no longer taken in Oxford but on the BM18 beamline of the European Synchrotron Radiation Facility (ESRF) in Grenoble, France — an instrument fine enough to separate the gossamer-thin, densely packed layers of a roll. The work was done in collaboration with the National Library of Naples "Vittorio Emanuele III," which safeguards the papyri.

From the X-ray volume, the software first reconstructs the geometry of the winding, traces the papyrus surface through the volume (segmentation), and flattens it into a readable page. Then the neural networks come in: trained on examples where one knows where ink sits, they learn to recognize and amplify the ink signal that is barely distinguishable from the carbonized papyrus. At the end there is always the human: every reading is examined, transcribed, and interpreted by papyrologists — the AI makes the signal visible, but judgment about meaning and language remains with expertise.

Crucially, all of this is open. The tomographic data, the reconstructed surfaces, the transcriptions, and the code are publicly available under a free license; the results can be read as a preprint on arXiv (arXiv:2606.29085). Anyone can check the work, build on it, and apply it to the hundreds of rolls that remain sealed. And, remarkably: most of today's research team originally arrived as contestants — people who won a prize and were then recruited.

A Framework: The Principles of Non-Invasive Reconstruction

What succeeded here follows a pattern that reaches far beyond carbonized papyri — into signal processing, machine learning, and the organization of research itself. It can be captured in a table.

The connecting thread is the separation of capturing, reconstructing, and interpreting. You capture, non-destructively, what is measurable; you reconstruct the hidden structure purely computationally; and only at the end does an expert interpret the result. Whoever mixes these steps — who believes, while still measuring, that they already know what should come out — spoils the result.

The Central Takeaway: The Signal Is There — Pull It Out of the Noise Without Destroying It

If you take a single transferable lesson from this story, let it be this: When information is physically present but hidden, the art lies not in grasping more firmly but in looking more finely — and in leaving the decisive step to a system that recognizes a signal invisible to your naked eye. For two hundred years, clever people failed because they wanted to open the rolls — that is, they did exactly what destroys the information. The breakthrough came when someone stopped grasping and started looking through.

This holds every time you stand before a black box whose interior you are not allowed to touch directly: a production system you cannot halt to debug; a trained model whose inner behavior you want to understand without retraining it; a dataset in which the pattern you seek drowns in the noise; a legacy system whose failure you cannot afford. The temptation is always to go in invasively — to stop, break open, alter. Resist it where you can. Ask instead: How can I observe the state without disturbing it? What weak, indirect signal tells me what is inside? And can the step from raw signal to meaning be handed to a learning method that recognizes finer differences than I can?

Concretely: before you next "open up" a sensitive system to get at some information, ask yourself whether there is a non-invasive path — a measurement that preserves the state, a structure that can be reconstructed purely computationally, a weak signal a model can amplify. And secure your result the way the team did: through a second, independent path that confirms the same answer.

Cross-References in the Vault

This story stretches several threads that run elsewhere in the vault:

• The task of recovering a lost meaning from pure structure links Herculaneum closely to the decipherment of Linear B: there as here, "structure before meaning" holds, only that this time the challenge is physical rather than linguistic.
• The loss of an entire body of ancient knowledge and its late recovery through reverse engineering shares its motif with the Antikythera Mechanism, that ancient computer whose blueprint was forgotten for over a thousand years.
• The library of Herculaneum comes from the world whose Bronze Age forerunner perished in the great Late Bronze Age collapse — another example of how fragile cultural knowledge is.
• The heart of the solution is ink detection by neural networks; anyone who wants to understand how modern AI learns from raw signals at all will find the foundations in the Transformer architecture and its image-generating relatives, the diffusion models.
• And the question of when a reading is really "knowledge" and not merely a lucky accident leads straight to the Gettier problem: the one-to-one confirmation from better data was exactly the gain in independent reliability that turns a justified belief into genuine knowledge.

A Closing Question for Reflection

For two hundred years, the mistake lay in wanting to open the rolls — that is, in the very act that destroys the information. Where in your own work do you reflexively reach in "invasively" — halting a system, breaking open a structure, altering a state — to get at information that might also be obtainable non-invasively, if you only looked more finely and left the decisive step to a method that recognizes weaker signals than your naked eye?

Sources

• "An entire Herculaneum scroll has been read for the first time" – Vesuvius Challenge (June 25, 2026): https://scrollprize.org/firstscroll
• "Complete virtual unwrapping and reading of a rolled Herculaneum papyrus" – Preprint, arXiv:2606.29085 (2026): https://arxiv.org/abs/2606.29085
• "New secrets revealed from the Herculaneum scrolls" – Diamond Light Source (2026): https://www.diamond.ac.uk/Home/News/LatestNews/2026/New-secrets-revealed-from-the-Herculaneum-scrolls.html
• "Herculaneum scrolls: A 20-year journey to read the unreadable" – EduceLab, University of Kentucky: https://educelab.engr.uky.edu/news/herculaneum-scrolls-20-year-journey-read-unreadable
• "Virtual unrolling and deciphering of Herculaneum papyri by X-ray phase-contrast tomography" – Scientific Reports (2016): https://www.nature.com/articles/srep27227
• "Title and author of burned, still-rolled scroll decoded after nearly 2,000 years" – CNN (May 6, 2025): https://www.cnn.com/2025/05/06/science/herculaneum-scroll-title-author-decoded-intl-scli
• "Inside the AI Competition That Decoded an Ancient Herculaneum Scroll" – Scientific American: https://www.scientificamerican.com/article/inside-the-ai-competition-that-decoded-an-ancient-scroll-and-changed/

← All articles