A computer scientist has labored for 21 years to read carbonized ancient scrolls that are too brittle to open. His efforts stand at last on the brink of unlocking a vast new vista into the world of ancient Greece and Rome.
Brent Seales, of the University of Kentucky, has developed methods for scanning scrolls with computer tomography, unwrapping them virtually with computer software, and visualizing the ink with artificial intelligence. Building on his methods, contestants recently vied for a $700,000 prize to generate readable sections of a scroll from Herculaneum, the Roman town buried in hot volcanic mud from the eruption of Vesuvius in 79 A.D.
The last 15 columns—about 5 percent—of the unwrapped scroll can now be read and are being translated by a team of classical scholars. Their work is hard, as many words are missing and many letters are too faint to be read. “I have a translation but I’m not happy with it,” says a member of the team, Richard Janko of the University of Michigan. The scholars recently spent a session debating whether a letter in the ancient Greek manuscript was an omicron or a pi.
The scholars believe that the manuscript is a work of Philodemus, a tutor of the great Roman poet Virgil and the house philosopher at the villa in Herculaneum where the scrolls were found. They can’t be sure because the author and title, usually written at the outer end of the scroll where they are working, have not been provided to them. Some suspect that this information is being withheld to test the validity of their attribution. Seales says this isn’t the case; his group is still looking for the author and title lines.
The artificial intelligence teams working on the ink are making rapid progress and hope that 85 percent of the first scroll will be visualized by the end of this year. Such an achievement would be extraordinary given the dismal history of attempts to read the scrolls. Some 800 were recovered during excavations at Herculaneum between 1752 and 1753. But repeated attempts to open the scrolls by various methods have destroyed most of them, yielding a few fragments of readable text. By the time scholars realized that the best course was to leave them alone, a mere 270 were left. These are the next scrolls to which the Seales team will turn its attention if given permission by the Italian authorities.
Beyond that may lay a far greater prize: the main library of the villa.
A magnificent structure that is the model for the Getty Museum in Los Angeles, the villa is believed to have belonged to a powerful Roman senator, Lucius Calpurnius Piso Caesoninus, whose daughter Calpurnia was the third wife of Julius Caesar. Most of the scrolls found so far come from a room thought to have been the private library of Philodemus and seem mostly to be technical works of Epicurean philosophy. But Roman aristocrats like Piso often owned large and important libraries, according to Cicero’s letters. Piso’s villa is known to have had a lecture room, suggesting a library nearby. The villa has been only partially excavated, mostly by tunnels dug in the eighteenth century. Two unexplored stories were discovered in the 1990s, and these could contain a main library with thousands of scrolls. Italian authorities so far have focused more on conserving Piso’s villa than on doing further excavation. But their plans could change if Seales’s scroll-reading project gathers momentum.
Seales has had to overcome daunting obstacles to reach this point, not all of them technical. The Italian authorities declined to make any of the scrolls available, especially to a lone computer scientist with no standing in the field. Seales realized that he had to build a coalition of papyrologists and conservationists to acquire the necessary standing to gain access to the scrolls. He was eventually able to x-ray a Herculaneum scroll in Paris, one of six that had been given to Napoleon. To find an x-ray source powerful enough to image the scroll without heating it, he had to buy time on the Diamond particle accelerator at Harwell, England.
In 2009, his x-rays showed for the first time the internal structure of a scroll, a daunting topography of a once-flat surface tugged and twisted in every direction. Then came the task of writing software that would trace the crumpled spiral of the scroll, follow its warped path around the central axis, assign each segment to its right position on the papyrus strip, and virtually flatten the entire surface. But this prodigious labor only brought to light a more formidable problem: no letters were visible on the x-rayed surface.
Seales and his colleagues achieved their first notable success in 2016, not with Napoleon’s Herculaneum scroll but with a small, charred fragment from a synagogue at the En-Gedi excavation site on the shore of the Dead Sea. Virtually unwrapped by the Seales software, the En-Gedi scroll turned out to contain the first two chapters of Leviticus. The text was identical to that of the Masoretic text, the authoritative version of the Hebrew Bible—and, at nearly 2,000 years old, its earliest instance.
The ink used by the Hebrew scribes was presumably laden with metal, and the letters stood out clearly against their parchment background. But the Herculaneum scroll was proving far harder to read. Its ink is carbon-based and almost impossible for x-rays to distinguish from the carbonized papyrus on which it is written. The Seales team developed machine-learning programs—a type of artificial intelligence—that scanned the unrolled surface looking for patterns that might relate to letters. It was here that Seales found use for the fragments from scrolls that earlier scholars had destroyed in trying to open them. The machine-learning programs were trained to compare a fragment holding written text with an x-ray scan of the same fragment, so that from the statistical properties of the papyrus fibers they could estimate the probability of the presence of ink.
Seales was then approached by Nat Friedman, former chief executive of GitHub, a platform for software developers. Friedman had taken an interest in the Herculaneum papyri after reading about them. He suggested involving the community of software engineers who work on artificial intelligence through a series of benchmark prizes that would foster both competition and the sharing of programs.
Scholars who develop a new technique or discover a new source will usually keep it to themselves until they have garnered its first fruits. “The coin of the realm in academia is credit. I thought long and hard about it,” Seales said of Friedman’s proposal for collaboration. He decided that Silicon Valley could scale up his project far faster than he could do with his small university research team.
“Brent decided to share the glory with anyone who showed up,” Friedman said. “He told me he had been working on it for 18 years and just wanted to see it through.”
Friedman raised more than $1 million for the incentive prizes and hired a team to augment, and if possible, automate, the scroll-unwrapping problem. A significant advance came last year, when Casey Handmer, one of the contestants, noticed a feature of the x-ray scans that looked like cracked mud and which he called “crackle.” The crackle turned out to be correlated with letters in the scroll. Because no one quite knows what the machine-learning programs have taught themselves, it’s hard to say if the crackle signifies real ink, or just the pressure of the copyist’s pen on the papyrus surface, or some combination of features. But no matter—Greek letters started miraculously to appear for the first time from the inchoate surface of the scanned scrolls.
Last August, the first of the new awards, called Vesuvius prizes, was awarded to Luke Farritor, a 21-year old student at the University of Nebraska. He received $40,000 for visualizing the first word to emerge from the scroll—πορφυρας, the Greek word for “purple.” The next competition was for the first person to visualize four passages, each of 140 letters. A $700,000 prize was awarded on February 4 to a team consisting of Farritor; Youssef Nader, an Egyptian Ph.D. student in Berlin; and Julian Schilliger, a robotics student from ETH Zürich.
The decoders of the scroll have raced so far beyond this requirement that the last 15 columns of the scroll can now be read for the first time in 2,000 years. The subject, as might be expected from a work of Epicurean philosophy, is pleasure, but gaps and other problems at present thwart a full translation.
“This proves that scrolls in this kind of state can be successfully deciphered without opening them—it’s astonishing,” Janko says. Many hope that progress with Napoleon’s Herculaneum scroll may inspire the Italian authorities to resume excavations at Piso’s grand villa. “There’s a moral imperative to go dig out that villa,” says Friedman. “It could be the greatest archaeological treasure on earth.”
If the main library exists, Janko would like to see the full poems of Sappho, regarded as the greatest lyric poet of her time but now known mostly through fragments of her work. Seales hopes to find documents from the early Christian era. St. Paul, after all, wrote an epistle to the Romans; Piso’s library might contain some reference to the young sect.
Janko reckons that maybe 2 percent of classical Greek and Roman literature has survived. Recovering an ancient library would vastly expand knowledge of the ancient world. “I think it’s more important than ever,” he says, “to find these books before Vesuvius erupts again.”
Photo: Vesuvius Challenge