AI Role in Rediscovering Lost Classical Texts

The artifact referred to as P.Herc.Paris.3 bears a striking resemblance to a dark, ash-gray mass, roughly akin in size and contour to a banana. This association has earned it the moniker “Banana Boy.” This intriguing object is, in reality, a papyrus scroll that was uncovered amidst the ruins of a villa situated in the ancient Roman city of Herculaneum, nestled within the region of Campania. This villa, along with its vast library of numerous scrolls, endured a process of carbonization when searing gases enveloped the city during the catastrophic eruption of Mount Vesuvius in 79 AD, which also entombed the nearby town of Pompeii.

While the scrolls managed to survive, their carbonized state has rendered the traditional unrolling procedure nearly impossible. Fast forward nearly two millennia, and a breakthrough has occurred. Words concealed within Banana Boy have now been unveiled for the very first time, thanks to the efforts of volunteers participating in a prestigious competition. Utilizing a combination of cutting-edge X-ray technology and artificial intelligence, they successfully virtually unfurled the ancient scroll.

The inaugural word to be uncovered, announced on October 12th, was “porphyras,” signifying “purple” in the ancient Greek language (as depicted in the image below). This remarkable achievement was accomplished by Luke Farritor, a computer science student hailing from the University of Nebraska-Lincoln, thereby earning him a well-deserved $40,000 prize. Mr. Farritor’s accomplishment built upon the prior work of Casey Handmer, a former NASA physicist, who meticulously examined X-ray images of Banana Boy’s charred layers. Handmer’s analysis revealed a distinctive “crackle pattern,” providing crucial clues to the presence of ink within the scroll.

Subsequently, Youssef Nader, a robotics student at the Free University of Berlin, successfully rediscovered the same word. (Both Dr. Handmer and Mr. Nader received $10,000 prizes for their achievements.)

Mr. Nader has since unveiled an image extracted from the scroll, displaying four columns of text placed side by side. For scholars of antiquity, this is an exciting development. It is believed that the villa in question was owned by Lucius Calpurnius Piso, who happened to be Julius Caesar’s father-in-law. The ability to decipher the contents of its well-stocked library could potentially expand the corpus of surviving texts from ancient times. Already, there is a buzz of anticipation regarding the possibility of forgotten plays, new philosophical works, or even lost Homeric poems.

Efforts to decipher these scrolls began in the 1750s when the villa was rediscovered. Early attempts to unroll them with knives only led to their disintegration. Recognizing their fragility, Antonio Piaggio, a conservator from the Vatican, constructed a machine in 1754 to slowly unwind them using weights and strings. Even so, the unrolled scrolls disintegrated, and the resulting fragments were extremely challenging to read due to the charcoal-based ink’s poor visibility against the shiny black charred papyrus. However, some characters that could be deciphered revealed that some scrolls contained philosophical works in ancient Greek.

Fast forward a quarter of a millennium to 1999, when scientists from Brigham Young University employed infrared light to illuminate some of these fragments. This technique created a stark contrast between the papyrus and the ink, making the writing more legible. Subsequent advancements included multi-spectral imaging in 2008, which used various wavelengths of light to reveal previously illegible words.

Many fragments were discovered to belong to texts written by the Greek philosopher Philodemus of Gadara, who had been previously known only through references in other works (with Cicero being a notable fan of his poetry).

Approximately 500 scrolls still remain unopened. Physical unrolling is no longer attempted due to the damage it causes. Instead, the focus has shifted to finding ways to virtually unwrap them using 3D scans of the rolled-up scrolls to generate a series of legible 2D images. A pioneer in this approach is W. Brent Seales, a computer scientist at the University of Kentucky, who, in 2009, arranged for the scanning of Banana Boy and another scroll known as Fat Bastard in a computerized tomography (CT) X-ray machine, typically used for medical scans. This produced detailed internal structure images but failed to make the ink within the scrolls discernible.

In 2015, Dr. Seales analyzed a different carbonized scroll discovered in 1970 at En-Gedi, near the Dead Sea in Israel. It had been written using a metal-rich ink, which was clearly visible in X-ray images against the papyrus (the text turned out to be the Book of Leviticus). This confirmed that, under the right conditions, digitally unrolling a carbonized scroll and reading its contents could indeed be achieved.

The next step was to amalgamate these existing methods into a novel approach. In 2019, Dr. Seales arranged for high-resolution scanning of Banana Boy, Fat Bastard, and four fragments of other scrolls using the Diamond Light Source in Britain, a particle accelerator capable of producing more powerful X-ray light than a CT scanner. He then combined infrared images, where ink is readily visible, with X-ray scans of the same fragments, where the ink is not visible.

Earlier this year, Stephen Parsons, a graduate student collaborating with Dr. Seales, utilized machine learning to train a model to recognize faint signs of ink in X-ray images using the infrared scans. This process, in theory, brought deciphering the scrolls down to a complex software problem. However, further refinement and scaling were required for this software.

Enter Nat Friedman, a technology executive and ancient Rome enthusiast, who offered to support Dr. Seales’s work. They decided to accelerate progress by organizing a contest, the Vesuvius Challenge, with a prize fund of $1 million. More than 1,200 teams entered, participating in various challenges aimed at improving ink detection and “segmentation,” the process of transforming 3D scans into 2D scroll surface images. Scrutinizing segmented images from Banana Boy, Dr. Handmer recognized that the crackle pattern indicated the presence of ink. Mr. Farritor harnessed this insight to refine a machine learning model, ultimately revealing legible words.

Meanwhile, Mr. Nader pursued a distinct approach. He initiated “unsupervised pretraining” on the segmented images, tasking a machine learning system to identify patterns without external hints. He fine-tuned the model using the winning entries from the Kaggle ink-detection challenge. After witnessing Mr. Farritor’s initial results, he applied this model to the same segment of Banana Boy, identifying some letters. Through iterations, his model improved in finding more letters. Papyrologists assessed all results before awarding the prizes.

Equally as significant as the technology itself is the innovative approach employed in this endeavor. It effectively applies the open-source software development methodology, an area of expertise for Mr. Friedman, to the realm of archaeology. “It’s a distinctive collaboration between tech entrepreneurs and academics, leveraging future-oriented tools to bring the past into the present,” he affirms. Dr. Seales estimates that the competitive aspect has compressed a decade’s worth of research into a mere three months.

A vibrant community of volunteers has now adopted these new tools to examine the two scanned scrolls. Mr. Friedman is optimistic, with a 75% chance, that someone will secure the grand prize of $700,000 for identifying four distinct passages of at least 140 characters by year’s end. “It’s a race now,” he proclaims. “We’ll be delving into entire books next year.”

Unlocking the secrets of Banana Boy represents just the tip of the iceberg. Only a fraction of ancient Greek and Roman literature has endured the passage of time. However, if the hundreds of other scrolls salvaged from the villa can be similarly scanned and interpreted, it has the potential to vastly enrich our repository of texts from antiquity. Dr. Seales holds out hope that the Herculaneum scrolls might harbor “an entirely new, hitherto undiscovered text.” Meanwhile, Mr. Friedman harbors a particular desire for the resurrection of one of the lost Homeric epic poems.

Furthermore, all of this could potentially rekindle enthusiasm for more comprehensive excavations of the villa, Mr. Friedman suggests. The scrolls presently recovered were found in a single corner of what scholars believe to be a considerably larger library spread across multiple floors. If this supposition holds true, it may contain thousands of scrolls in Greek and Latin.

It’s worth noting that one reason classical texts are so rare is the limited durability of the papyrus on which they were originally written, particularly in Europe’s temperate, rainy climate. Dr. Seales finds it poetically ironic that the carbonization of the scrolls, which complicates their readability, is also what has preserved them for posterity. Furthermore, fragments of scrolls that disintegrated during physical unrolling have ultimately provided the key to unrolling the remainder virtually.