AI helps historians illuminate the past
An AI model developed at KAUST reconstructs inscriptions on ancient statues, helping historians analyze sources that might otherwise remain unreadable.
The model, named DeepHadad, is trained to reconstruct damaged inscriptions to make them much easier to read.
Unlike AI tools that use language patterns to predict missing words or letters, DeepHadad analyses the physical structure of inscriptions.
To train the model, the researchers simulated damage to inscriptions, creating paired examples of intact and damaged texts that are scarce in existing sources.
What is DeepHadad and what does it do?
Inscriptions on statues are a vital resource for historians trying to understand ancient civilizations. However, weathering, erosion and intentional damage have caused much of this written record to be lost over the centuries.
Researchers at KAUST have developed DeepHadad, an AI model that creates digital reconstructions of physical surfaces that have deteriorated over time. The team has used the technology to improve the legibility of the heavily damaged inscriptions on an almost 3,000-year-old Hadad statue.
“DeepHadad helps historians and archaeologists study ancient written records that might otherwise be impossible to decipher,” says Jonathan Klein, KAUST research scientist who worked on the project. “This breakthrough opens new possibilities to enrich our understanding of the past.”
Inscriptions on the Hadad statue before and after reconstruction using DeepHadad. © KAUST 2026.
How was DeepHadad created?
The researchers first used 3D scanning to capture the surface of the inscriptions on the Hadad statue, recording the depth and shape of the carved letters. This produced detailed digital representations that the model could use to learn the physical features of the inscriptions.
A major challenge was the lack of training data. AI models are typically trained on vast datasets containing trillions of data points. However, there are only a limited number of surviving inscriptions in Aramaic, the ancient Northwest Semitic language used on the Hadad statue. The researchers also faced the difficulty that each carved letter has a unique shape, meaning they could not find matching examples of the same letter in both intact and damaged states.
By combining computational methods with historical expertise, we can advance the study of ancient inscriptions and gain a deeper understanding of the past.”
Dominik L. Michels
To overcome this limitation, the team used well-preserved letters from the statue and digitally simulated damage such as erosion, cracks, and deformation. This generated thousands of paired examples of intact and damaged letters, giving DeepHadad the training material it needed to learn how to reconstruct damaged inscriptions and make them easier to read.
How can AI support historians and archaeologists without replacing them?
DeepHadad is a valuable new tool for scholars, not a replacement for them. The model can make inscriptions much easier to read, but it cannot necessarily recover every missing or damaged letter.
Furthermore, DeepHadad does not try to understand or predict what each inscription says. Its role is to make damaged text more legible, leaving historians and archeologists to interpret its meaning and place it in its historical context.
“Even with extensive training data, interpreting ancient languages requires the expertise of human specialists,” says Dominik Michels, KAUST professor of computer science. “By combining computational methods with historical expertise, we can advance the study of ancient inscriptions and gain a deeper understanding of the past.”
REFERENCE:
