Reading All Herculaneum Scrolls
One scroll is read in full, but hundreds stay sealed and much of the villa that held the library is still buried.
Open in the interactive tree →The Villa dei Papiri yielded about 1,800 papyri fragments and rolls in the 1750s, mostly Epicurean philosophy; many that were opened by hand were damaged. Reading the sealed rest needs a scan of each scroll, human guidance that the team says is still needed, and papyrologists to edit the text. Scholars argue that lower levels of the villa hold more of the library, and a 2016 report found no new digging since 2009.
As of October 2026
scrollprize.org lists 45 scrolls and fragments scanned, ink found in 9, and one scroll read end to end (PHerc. 1667, announced 25 June 2026). The 2027 Grand Prize of $1M asks for another full scroll with at most eight hours of human input; of a $2.14 million prize pool $1.868 million has been awarded. The ESRF puts the unopened scrolls at more than 600 (counts vary between sources) and says its scans run to up to 300 TB per scroll, an upper bound near 180 petabytes for all of them. No restart of excavation appears in the sources found.
What is missing
- Unwrapping with human input cut to a few hours or less per scroll
- Ink detection that works on scrolls where the ink is faint or absent
- High-resolution scans of all sealed scrolls and the storage to hold them
- Enough papyrologists to edit and publish the texts
- A decision and funding to excavate the unexcavated levels of the villa
Becomes possible once solved
- Lost works of Epicurean and other Greek philosophy
- A fuller picture of one Roman library
- Methods for other sealed or burned manuscripts
Open steps
- Petabytes of scan data Medium AI leverageAt the ESRF's count of 600+ sealed scrolls and up to 300 TB each, scans would approach 180 PB (upper bound, by arithmetic); storage and processing need planning.
- Which scrolls hold readable ink? Medium AI leverageInk shows in 9 of 45 scanned scrolls and fragments and the minimum resolution to see it is unknown; learn which carry a signal before spending beamtime.
- Enough papyrologists Low AI leverageEach scroll's text must be edited, dated and attributed; the ESRF says the focus is shifting to papyrologists, classicists and historians.
- Excavate the rest of the villa Low AI leverageMuch of the villa is buried and lower levels were only partly explored; a 2016 report found no new digging since 2009, with cost and site stability as obstacles.
Where AI could help
High AI leverage. Unwrapping, ink detection and letter reading are image problems that machine learning already speeds, and the 2027 prize caps human input at eight hours.
- Trace papyrus layers automatically across hundreds of scrolls
- Detect and verify ink with independent models
- Draft transcriptions and restorations for scholars to correct
- Decide which scrolls to scan first
Shown so far
- The Vesuvius Challenge's new $1M Grand Prize needs another full scroll with at most eight hours of human input; the team says full automation is required to read the collection in a reasonable time. source
- The project's pipeline scans scrolls, flattens their layers, and uses machine-learning models to find ink; it lists hallucinated letters and an unknown minimum scan resolution for ink as open difficulties. source