A University of Iowa-led research team is aiming to use AI to restore lost sections of ancient manuscripts, with support from a $500,000 grant from Schmidt Sciences, a philanthropic organization that funds unconventional scientific research.

Paul Dilley, the Erling B. “Jack” Holtsmark Associate Professor in the UI’s Classics department and the departmental executive officer of the UI Department of Religious Studies, will train AI models for the next three years to suggest hidden characters in ancient texts.

Dilley said many ancient texts are uncovered by happenstance, such as the Herculaneum papyri, a collection of thousands of ancient Roman scrolls carbonized by the A.D. 79 eruption of Mount Vesuvius.

“Often these texts aren’t very well preserved,” he said. “You need to take steps to make them legible again. Sometimes, those steps are difficult, as in the case of the Herculaneum papyri.”

Many ancient texts are filled with ink that has faded to the point where the naked eye can no longer discern it.

“The first step in recovering these texts is usually some kind of enhanced imaging technique to try to make the ink appear more legible,” Dilley said. “Multispectral imaging can make the traces of ink that are still left on the page really jump out.”

Dilley said he and his team will use multispectral imaging as a first step to identify ancient manuscripts, then bring AI to bridge any gaps.

Honghai Zhang, a senior research scientist at the Iowa Institute for Biomedical Imaging — a multidisciplinary research initiative at UI that brings together different departments for medical imaging — used a similar form of multispectral imaging in a study analyzing medieval manuscripts.

Zhang was part of a team that used medical CT scans, which use X-rays to take detailed images of the body, to look inside the books without opening them.

“To extract hidden information from a damaged manuscript, parchment, or scroll, it’s a huge effort,” he said. “It involves the collaboration of people from different research fields.”

Zhang said the chemical buildup of the ink in a manuscript will play a large role in determining which scanner will best decipher the text. While Zhang’s team used CT scans, others are using infrared-based scans, a process that takes several hours.

“Imaging has been evolving and advancing for many years. Day in and day out, we’re dealing with higher and higher resolutions,” he said. “AI is also evolving, but using the combination of those two for ancient manuscripts is something new, which is interesting and challenging.”

RELATED: UI study reveals Africa’s ancient apex predator

Dilley said even as multispectral imaging continues to advance, there is a limit to what it can decipher.

“If the manuscript has already been damaged, imaging obviously doesn’t help you see the missing parts,” he said. “That’s where AI comes in.”

Dilley said editors will train the AI to perform a process called infilling, or suggesting restorations for letter fragments. Generative AI will also be used to infer and suggest new text for lacunas, or portions of manuscripts where all text is missing.

For his current study, Dilley said the AI will be trained to perform infilling and restore lacuna for Greek, Latin, and Celtic languages.

“The plan is to publish the models open access and to make it extendable to other languages,” he said. “The basic pipeline should be extendable to other languages.”

Robert Cargill, the Roger A. Hornsby associate professor in the classics, said he and Dilley were hired at UI as part of a Digital Humanities Initiative in 2011, an emerging field that had just begun digitizing ancient texts and putting them in online libraries.

Cargill said Dilley’s new experimentation with AI is a similar instance of having the foresight to use emerging technology for studying the humanities.

“There is some research that can be done without technology, and technology simply helps speed up the process,” he said. “But there is other research, like the kind that Paul is doing, that just cannot be done without the help of technology and AI.”

Cargill said opening something like an ancient charred scroll would be impossible; simply, as it would crumble from frailty. With an AI algorithm that scans every layer and compares lines to a database of known letters in the ancient language, Cargill said such work is possible.

“This is real research that 50 years ago we would have just called magic,” he said. 

Cargill said Dilley’s research, being conducted at Iowa, will help UI students learn about the emerging role of AI in the humanities, giving them a leg up in the job market. 

“The only other people doing it are nations. Israel is doing this with the Dead Sea Scrolls, but that’s an entire government,” he said. “Countries that have massive budgets are doing this, and Professor Dilley is able to do it here at the University of Iowa with this grant.”

Dilley said as his team looks to integrate AI into the humanities, it is important to give scholars space to have the final say on what lost sections of ancient manuscripts could have been.

“It’s definitely not meant to replace traditional human language learning and engagement, but it can be a way to make potential breakthroughs in otherwise very difficult sections of a damaged manuscript,” he said. “It’s meant to be a great help, but not the final say.”