Parametric reverse engineering assisted by artificial intelligence. The case of the roman imperial baths in North Africa

Naifer, Zaineb [UCL]

This research addresses the conservation of roman imperial baths in North Africa, which are confronted with a number of challenges. Some sites are well-preserved and complete, while others are in a state of ruin, with only partial plans available. These issues present significant complexities for archaeologists, architects, and other heritage professionals, who often rely on historical sources and previous interventions to reconstruct these buildings. In order to overcome these challenges, the thesis proposes a two-phase methodological approach. The initial phase employs parametric reverse engineering to delineate the methodologies and conventions utilized by roman architects, thereby generating a parametric data set model (PDSM). The second phase employs artificial intelligence (AI) to develop a generative model (IAGM). A multilayer neural network is trained with the PDSM and subsequently evaluated through the provision of plan fragments, thereby assessing its capacity to generate complete plans from incomplete data. The study examines 29 roman imperial baths (23 in North Africa and 6 in Rome), with varying degrees of excavation progress and document availability. The developed method employs a multimodal approach, beginning with the prediction of programs from fragments, the deduction of minimal programs, the delineation of missing components, and the assembly of spaces. In doing so, it considers a range of parameters, including graphic, geometric, and dimensional data. It is anticipated that the results will include models that can be used by practitioners, the creation of knowledge for students of architecture, history and computer science, and the ability to integrate new data in order to refine the results. Furthermore, this approach enables the conceptualisation of novel bath designs, thereby contributing to the preservation and reconstruction of roman architectural heritage.