Machine learning reads a buried Herculaneum scroll end to end

Researchers have read a Herculaneum scroll from end to end for the first time, recovering a text that has been sealed since Vesuvius buried it in 79 AD. The scroll, known as PHerc. 1667, was never physically unrolled. It was scanned at the European Synchrotron Radiation Facility, then read by machine learning models trained to spot ink that is almost invisible against the carbonized papyrus underneath it.

The result is concrete: roughly 1.4 meters of papyrus and about 22 columns of ancient Greek, a philosophical treatise on Stoic ethics. The team behind it grew out of the Vesuvius Challenge, an open competition where many of the researchers first showed up as contestants. That detail matters, because it means the hard part was not one lab's secret method but a problem a public contest could attack from many angles, with EduceLab under Brent Seales and the National Library of Naples providing the scans and ground truth.

Reading one scroll is the proof of method. Hundreds of carbonized scrolls from the same library remain unread, an entire collection of ancient philosophy and prose that no one has seen for two thousand years. The bottleneck now shifts from whether it can be done to how fast it can scale.

Why it matters

This is a clean example of machine learning recovering information no human eye can see, applied to a problem with a verifiable answer. If you build ML systems, the model here is worth studying: a public competition with shared data and a clear metric did what years of closed effort could not.