When Mount Vesuvius buried Pompeii in volcanic ash more than 2,000 years ago, it also buried a library of scrolls in the nearby town of Herculaneum that scholars have struggled to read since their discovery more than 250 years ago.
But Herculaneum, the only surviving library from the classical world, could soon reveal its hidden history, thanks to researchers at the University of Kentucky. They’re using artificial intelligence, GPUs and a computer imaging program they call “virtual unwrapping” to digitally separate and flatten the tightly rolled sheets so they can be read.
“There’s an Indiana Jones aspect to it,” said W. Brent Seales, chair of the University of Kentucky Computer Science Department, who has labored for more than 13 years on ways to read the scrolls.
Charred by the lava’s intense heat, the Herculaneum scrolls can easily be mistaken for oversized cigars or spent logs. To put it plainly, Seales said, “they look like hell.”
How Not to Unwrap a Scroll
The delicate papyrus scrolls can’t be physically unrolled without serious damage. About 20-30 feet long and tightly wound 100 times around, the scrolls are brittle and permanently creased. In each scroll, the layers of papyrus are welded together.
Separating one layer from another is something like trying to pull apart the flaky layers of pastry in croissant.
Early efforts to unwrap the scrolls were disastrous. Not long after the scrolls were discovered, a Vatican conservator invented a machine that successfully unwrapped some, but at a high price — it rendered one side of every scroll completely unreadable. Later attempts using substances like mercury or rose water damaged the parchment on which the text is inscribed. Others left pages stuck together, trapping hidden layers and their precious contents.
As late as 1985, scientists tried to loosen the wraps using a mixture of ethanol, glycerin and warm water. One scroll was peeled apart into many fragments. The other dried up and broke into more than 300 pieces.
They See Me Scrollin’: Ein Gedi Scroll
Seales and his team developed a way to unfurl the scrolls virtually, using a computed tomography scan — the kind of X-ray that produces detailed digital images of biological tissues. They cut the CT scan into vertical slices to show the internal structure of the scrolls. Then, the team’s software digitally flattens and reassembles the sheets so researchers can read them.
In September, Seales used this “virtual unwrapping” method to help identify a damaged scroll found in a synagogue in Ein Gedi, Israel, as the first two chapters of the biblical Book of Leviticus. (Read more about virtual unwrapping and the Ein Gedi scrolls in the team’s article in Science Advances.)
Reading the Scrolls: Black on Black
But the Ein Gedi find was a single scroll five layers thick. The Herculaneum library consists of hundreds of thick scrolls.
Before digitally unwrapping these more complex scrolls, the researchers had to be able to distinguish and separate the individual layers within. For this, they needed GPU-accelerated machine learning.
Although Seales and his team are still perfecting their methods, the goal is to teach the computer to automatically detect a layer and then accurately track each point along the layer until it has accurately identified a complete wrap. Text can be scrambled if the layers are mixed up, Seales said.
Even after that’s accomplished, a remaining obstacle is reading the text. Ancient inks were made of vegetable dyes or, often, burnt wood and oil and would have stood out on the papyrus if it wasn’t charred to a blackened crisp. The CT scan cannot distinguish between the burnt papyrus and the ink.
Although Seales used machine learning for his earlier work, he’s now testing deep learning to read the writing on the scrolls and more effectively separate the scrolls’ layers.
“The technical challenges are huge,” Seales said. “We’ve only now reached the place where companies like NVIDIA are building hardware that can help us solve the (computationally intensive) problems. That didn’t exist when I started working on this.”
In this case, the hardware is our GeForce GTX 1070 GPUs. For its deep learning tests, the team is using the TensorFlow deep learning framework, configured with cuDNN and our CUDA parallel computing platform.
Researchers are now working on about 250 scrolls; most believe the library contains some 1,800 scrolls in all. If AI unlocks the secrets of the Herculaneum scrolls, scholars hold out hope of recovering the lost works of Greek and Roman literature.
“If we perfect a method for reading these materials and the others are excavated, it would be the biggest find in 500 years,” he said.