Reading the Burned Library of Herculaneum

This is a Present piece, which means it is written to be rewritten. The state of play described here was last reviewed on the date above. Before republishing, check the latest results and update the section at the end.

When Vesuvius buried Herculaneum in 79 CE, it buried a private library along with the town. The scrolls inside were not burned to ash. They were carbonised, turned into brittle black cylinders by heat and then sealed under volcanic material.

They survived. They just could not be read.

Two Centuries of Destroying Them

The scrolls were rediscovered in the eighteenth century, and the early attempts to open them destroyed a great many. Unrolling a carbonised scroll physically is close to unrolling a cigar of charcoal. Some were sliced apart. Some were dissolved with chemicals. Fragments were recovered, and a portion of what was recovered turned out to be philosophical work associated with the Epicurean school.

Then the attempts stopped, because the remaining scrolls were too precious to keep breaking.

Reading Without Opening

The current approach does not touch the scroll at all.

A scroll is scanned with a particle accelerator, producing a high resolution three dimensional map of its internal structure. Software then has to do two hard things. It has to trace the individual wrapped layers through the volume and flatten them into virtual sheets, a process called segmentation. Then it has to find the ink.

The ink is the difficult part. The carbon based ink used sits on carbonised papyrus, so the writing and the page are close to the same material. What separates them is subtle, a slight difference in texture and density rather than a clear contrast. Machine learning models trained on fragments where the answer is already known have learned to pick out that difference.

The result is text appearing on a page that has not been physically opened in nearly two thousand years.

Why It Matters Beyond the Scrolls

Almost everything surviving from classical antiquity survived by being copied, repeatedly, by scribes who chose what was worth the effort. The Herculaneum library did not go through that filter. It is a snapshot of what one household actually owned, including work that later copyists ignored and which therefore vanished everywhere else.

If more of the library becomes readable, some of what appears will be text no one has read since the first century.

The method also generalises. Any object too fragile to open is a candidate, and there are a great many of those in museum storage.

State of Play

Early passages have been recovered and published, and the work has moved from proving the technique is possible to scaling it up across more scrolls. Progress depends on scanner time, on segmentation, which remains slow and partly manual, and on there being enough labelled training data.

There are also scrolls still in the ground. The villa has never been fully excavated, and the part of the library that has been read may not be the largest part of it.

To update this piece: revise the section above with the current count of scrolls opened, any new texts identified, and the status of further excavation at the villa. Change the updated date in the frontmatter when you do.