The Vesuvius Challenge

In 79 AD, the eruption of Mount Vesuvius buried the Roman town of Herculaneum, and with it a private library of papyrus scrolls. The heat didn't burn the scrolls to ash — it carbonized them, turning fragile paper into brittle, ink-black rolls of charcoal. Rediscovered centuries later, hundreds of these Herculaneum papyri are still rolled shut. Physically unrolling one tends to shatter it into fragments; the few that were forced open in the 18th and 19th centuries destroyed as much text as they revealed.

So the scrolls sat in museum drawers, unread, for 250 years. Then in 2023, entrepreneurs Nat Friedman and Daniel Gross teamed up with computer scientist Brent Seales — who had spent two decades scanning ancient manuscripts — to launch the Vesuvius Challenge: a set of cash prizes for reading a carbonized scroll without ever opening it.

The idea sounds like magic: pass the scroll through a CT scanner, reconstruct its layers in 3D, virtually "unroll" the spiral into a flat sheet, and train a model to spot ink that is nearly invisible even in the scan. Each of those steps is its own hard computational problem — and together they turned an archaeology puzzle into a machine learning one.

Below is a cross-section of a rolled-up "scroll" — a spiral, like the papyrus seen edge-on inside a CT scan. Hidden ink strokes sit on its surface, but coiled up like this they're impossible to read.

<p class="hint">{{hint_para}}</p>
<div class="stage">
  <canvas id="spiral" width="220" height="220"></canvas>
  <canvas id="strip" width="420" height="140"></canvas>
</div>
<div class="status" id="status">{{status_idle}}</div>
<div class="btns">
  <button id="trace" type="button">{{btn_trace}}</button>
  <button id="unroll" type="button">{{btn_unroll}}</button>
  <button id="reset" type="button" class="ghost">{{btn_reset}}</button>
</div>

* { box-sizing: border-box; }
body { font-family: system-ui, sans-serif; color: #222; margin: 0; }
.hint { font-size: .9rem; color: #444; margin: 0 0 .7rem; line-height: 1.45; }
.stage { display: flex; gap: .8rem; flex-wrap: wrap; align-items: flex-start; margin: .4rem 0; }
canvas { background: #f4efe4; border: 1px solid #d8cfba; border-radius: 8px; }
.status { font-size: 1rem; font-weight: 600; margin: .5rem 0; min-height: 1.4em; }
.status.ok { color: #0a7d33; }
.status.info { color: #1d3557; }
.btns { display: flex; gap: .5rem; flex-wrap: wrap; }
button { font: 600 14px system-ui, sans-serif; padding: .45rem .9rem; border: 1px solid #1d3557;
         background: #1d3557; color: #fff; border-radius: 8px; cursor: pointer; }
button.ghost { background: #fff; color: #1d3557; }
button:disabled { opacity: .5; cursor: default; }

// Code not found

Press Trace the layers to have the computer follow the spiral outward, sampling the surface as it goes — this is what a segmentation algorithm does to a real CT volume. Then press Unroll to lay that traced ribbon out flat. Only once the geometry is undone does the hidden text — or in the real Challenge, a machine learning model's ink predictions — become something you can actually read.

Nothing about this is a single algorithm — it's a pipeline where every stage can fail silently.

• Reconstruction. A CT scanner passes X-rays through the scroll thousands of times and mathematically inverts the attenuation into a 3D grid of densities. Papyrus and the air gaps between its coiled layers absorb X-rays almost identically, so the scroll's own geometry is barely visible in the raw scan — this is a much harder reconstruction regime than a hospital CT of bone and tissue.
• Segmentation. Software must trace each papyrus sheet through the coil, layer by layer, across thousands of scan slices, without ever losing the surface or crossing onto the neighboring layer. Done by hand this took teams of annotators weeks per scroll; it is the same kind of surface-tracking problem as watershed segmentation, just in 3D and at brutal scale.
• Virtual unwrapping. Once a layer is traced as a mesh, it gets algorithmically flattened into a 2D image — a texture-mapping problem, since a curved, crumpled sheet cannot be unrolled to a perfect rectangle without some distortion.
• Ink detection. Here is the real twist: carbon ink on carbonized papyrus is chemically almost the same material as the papyrus itself, so it is essentially invisible in a CT scan to the naked eye. The breakthrough was training a convolutional neural network to detect faint 3D texture patterns correlated with ink — using tiny scraps of scroll where ink truly is visible (via X-ray fluorescence at a particle accelerator) as ground-truth labels, then generalizing that signal to unopened scrolls where no ground truth exists at all.

Put together, this is why the $700,000 Grand Prize — first readable passages from an unopened scroll — wasn't claimed until October 2024, by a team led by Youssef Nader, Luke Farritor and Julian Schilliger. A further $200,000 prize for reading an entire scroll's contents remains open as of this writing: more text, less certainty, and machine-learning noise compounding across every unread layer.

"See inside something without touching it, then make the invisible legible" turns out to be a widely useful shape for a problem:

• Cultural heritage. The same pipeline is already being pointed at other fragile, unopenable objects — damaged medieval bindings, fire-damaged manuscripts, and other scrolls from the same Herculaneum villa still sealed in museum vaults.
• Medical imaging. CT and MRI reconstruction from noisy, indirect measurements is the same inverse problem as scanning a scroll, and machine-learning-assisted segmentation now routinely traces organs and tumors through 3D scan volumes.
• Industrial and forensic inspection. Detecting a faint, buried signal — a hairline crack inside a turbine blade, ink under a scorched document — is the same weak-signal classification challenge as finding carbon ink inside carbon papyrus.
• Crowdsourced, prize-driven research. The Challenge's open data and public leaderboard turned a single lab's multi-year project into a global competition that produced a winning technique in under two years — a template other stalled scientific problems have started to copy.

Learn how the Vesuvius Challenge works and you've touched non-invasive 3D reconstruction, surface segmentation, and the machine-learning art of pulling a signal out of near-total noise.

For 2,000 years, the only way to read a Herculaneum scroll was to destroy it. The Vesuvius Challenge replaced that trade-off with a computational one: scan it, trace its surface through a 3D fog, flatten the geometry, and let a model trained on scraps of visible ink find the invisible kind. None of these four steps is fully solved in general — each is its own open research problem — but together, for the first time, they produced real sentences from a scroll that had not been touched since Vesuvius sealed it shut.

The $700,000 Grand Prize is gone; the $200,000 prize for a fully-read scroll is still on the table. Somewhere in a museum drawer, an unopened library is waiting on the next model good enough to see through it.