Reconstructing the ancient world has always meant working with fragments. What survives today is often little more than scattered ruins, while surviving texts tend to mention buildings in passing rather than describing them in any real detail. Filling that gap, figuring out what a structure actually looked like from a handful of stones and a few offhand references, has traditionally taken years of careful scholarly judgment, piecing together incomplete evidence one inference at a time.
A team at Georgia Tech is now asking whether AI can take on part of that work, not by generating pretty pictures, but by learning the actual rules ancient builders followed.
The site
The study examines Athens’ Ancient Agora, a major center of political and social activity that played an important role in the development of democracy. Archaeologists from the American School of Classical Studies have excavated the site for nearly a century, uncovering more than 180,000 artifacts and recording them digitally. This extensive collection makes the Agora a valuable source for research, but architectural historian Myrsini Mamoli notes that its sheer size makes it difficult for any single researcher to analyze all of the material by hand.
Mamoli teaches both architecture and history at Georgia Tech and has spent years conducting fieldwork at the Agora. Her research includes Section Iota, an area that underwent major changes over the centuries. People moved artifacts and architectural fragments from other parts of the site into this section and later reused them in new structures, which makes their original locations difficult to identify. Archaeologists must therefore spend considerable time comparing physical evidence, excavation records, and historical context to determine where each fragment originally belonged.
The idea
The project draws on a concept from linguistics that treats architecture as a rule-based system, much like a language. Known as shape grammar, this approach suggests that architectural styles follow recognizable principles that determine how different elements fit together. In classical Greek architecture, designers used recurring proportions, spatial layouts, and structural features, combining them in consistent patterns to create coherent buildings.
Mamoli previously applied this method herself during her doctoral research. She painstakingly calculated relationships between architectural dimensions and used those measurements to create two dimensional reconstructions. Although the technique produced useful results, it required a huge amount of repetitive manual work. That experience raised a bigger possibility for her: instead of having researchers calculate every architectural relationship themselves, perhaps a computational model could identify those recurring rules automatically by studying a large collection of existing data.
AI’s role
This is where Kartik Goyal’s expertise becomes important. As an assistant professor, Goyal works at the intersection of natural language processing and humanities research. His earlier projects have involved using computational methods to examine early modern English books, including identifying their places of production by detecting small but meaningful patterns in their language and typography. After hearing Mamoli explain the architectural reconstruction problem, he recognized a similar computational task: using incomplete evidence to uncover the hidden patterns that connect separate pieces of information.
The team therefore wants to develop models that can understand the structure behind an architectural reconstruction rather than simply create an image that appears convincing. Goyal’s aim is for the system to work with geometric relationships and design principles themselves, instead of treating reconstruction as a surface-level exercise in generating pixels. Such a system could potentially combine the same kinds of evidence historians already rely on, including surviving fragments, excavation documentation, and descriptions from ancient writings, and use them to produce several possible versions of a lost structure. Each proposed reconstruction would be constrained by the architectural principles and recurring design patterns that likely guided the original builders.
Limitations
The researchers are also careful to acknowledge the limitations and dangers of the technology. Goyal notes that AI systems can have difficulty recognizing cultural context and may generate false or biased information while presenting it with complete confidence. That problem is especially serious here because the system would be making educated guesses about parts of history where the original evidence is already incomplete. For that reason, Goyal emphasizes the importance of creating models whose reasoning can be examined, questioned, and revised rather than systems that function as unquestionable black boxes.
Mamoli draws a similar line when describing the role of AI in the research. She views the technology as a tool that can support historians, not one that can take over their work or replace their expertise. In her view, interpreting archaeological evidence still requires the judgment of specialists who understand the historical and architectural context. At the same time, archaeological excavations continue to produce enormous amounts of information, making it increasingly difficult for researchers to process everything manually. Computational tools, she suggests, could eventually become an important way of managing that growing workload without removing humans from the interpretive process.
Current status
It is important to separate the project’s goals from what the researchers have actually achieved so far. The researchers secured $225,000 from the National Endowment for the Humanities earlier this year, but they are still developing the AI models. The team has not yet published a study with results, completed an independent evaluation, or released a dataset or benchmark for testing the system’s accuracy. For now, the project represents a promising research direction built around a particularly rich archaeological site.
The potential application is nevertheless quite significant. Mamoli points to a practical example involving fragments discovered in a disrupted area such as Section Iota. A successful system could potentially examine those isolated pieces and determine where they originally belonged within the Agora, even when they were found far from their original locations. That would move AI beyond simply producing visually convincing images. Instead, it could help researchers solve complex archaeological puzzles by connecting scattered evidence and reconstructing the history of a site on a scale that would be extremely difficult to manage manually.
