Thanks to machine learning, archeology has made leaps forward. A method has been developed using AI to read burned or otherwise damaged papyrus scrolls. There are many such documents and more and more are being found. The technology still needs to be refined, but the first results have been successful.
Back in the 18th century, excavations at a Roman villa in Herculaneum uncovered more than 1,000 whole or partial scrolls in a mansion believed to have belonged to Julius Caesar’s father-in-law. Eruption of Vesuvius in 79 AD. and the subsequent burial of the scrolls in the earth turned them into charred remains, which could only be unrolled once – at the same time they crumbled. The texts were also unreadable because the ink had faded along with the base. Reading all this and more is worth the effort.
Efforts to create technology to read the charred scrolls from Herculaneum were led for many years by University of Kentucky computer scientist Brent Seales. He and his team, using X-ray tomography at the Diamond Light Source accelerator complex, a third-generation synchrotron radiation source in Oxfordshire, learned to recognize traces of ink in papyrus fibers without touching or destroying the scroll.
But recognizing ink is just the beginning. It was necessary to “unroll” the scroll and read the text. To do this, in 2023, with sponsors’ money, the Vesuvius Challenge competition was announced with a prize of about $1 million. By the end of the year, leaders began to be identified. In particular, Luke Farritor, a computer science student at the University of Nebraska at Lincoln, was declared the winner of the “First Letters” stage for deciphering the first coherent phrases from the burned text, for which he received $40 thousand.
Farritor was later joined by Youssef Nader from Germany and Julian Schilliger from Switzerland. They developed an algorithm for “unfolding” scrolls. Together they were able to read more than 2,000 letters from the scroll. How became known the other day, a prize of $700 thousand went to this team.
Artificial intelligence copes with the task in several stages. The scroll is divided into sectors with each layer defined. Several methods have been proposed to solve this puzzle. For example, AI tracks the web of cracks in each layer, allowing it to pinpoint the layer and then align its digital copy. This is perhaps the most difficult part of the job.
Recognition of Greek letters also does not come directly from text, which is important for validating the team’s experience with outside research teams. All data is taken from a database obtained by X-ray tomography, and not using optical character recognition programs. It is also important that the code proposed by the group of winners of the competition is open and can be used by other groups to check the results. And they were confirmed. The AI actually reconstructs the text from detected ink residues in the fibers of the scrolls.
The technology is far from perfect, but its capabilities promise to bring many new things to our knowledge of the past. It can also be applied to reading the texts of the papyri in which mummies were wrapped. There are piles of these papyri in every decent museum, and they are a treasure trove of information about life thousands of years ago.