AI-powered built heritage: enhancing interpretation and recovery processes with generative AI models

Effective intervention and valorization activities of built heritage require robust knowledge representation and management, yet traditional methods often struggle with the complexities of diverse data sources and evolving knowledge structure. This paper presents an approach leveraging AI-driven methodologies, specifically Large Language Models (LLMs) to improve interpretation and recovery efforts in heritage preservation. The proposed framework introduces a novel methodology utilizing Generative Pre-trained Transformer (GPT) to facilitate knowledge extraction and interpretation within the heritage practices as the objective of the research. The approach encompasses phases including data collection, preparation and prompt engineering to train the tailored AI model with specific domain knowledge. The application to the Sanctuary of Hercules and the former Segrè papermill, shows the experiment results of the model trained through customization of LLM and evaluated through a qualitative prompt-response process. The results show that text-based data can be processed by the model quite effectively. In the meantime, image-based information is far more effective in the learning stages if it is enhanced with labels and metadata tagging. By bridging the gap between traditional methodologies and AI-driven technologies, the impact of the research lies in that the developed workflow enhances data accessibility, interpretation accuracy, and supports informed decision-making processes in built heritage. The paper contributes to advancing the integration of AI in heritage preservation, offering valuable insights for further research and practical applications in the field.