Machine Learning in Architectural Heritage Conservation: A Systematic Review and Thematic Synthesis of Applications, Challenges, and Future Directions

Machine learning (ML) is increasingly used in architectural cultural heritage conservation, but existing evidence remains fragmented across tasks, data types, and technical workflows. This systematic review synthesizes 33 studies on machine learning applications in architectural heritage conservation and identifies major application domains, methodological patterns, evidence gaps, and future research directions. Following a PRISMA-oriented review process, the literature was searched in Web of Science, Scopus, IEEE Xplore, ScienceDirect, SpringerLink, and Google Scholar up to 31 January 2026. Eligible studies were screened according to predefined inclusion and exclusion criteria, extracted using a structured coding form, and synthesized through qualitative thematic analysis. The included studies were grouped into three major domains: semantic understanding of heritage data, multi-scale damage detection and structural health monitoring, and system-level integration through heritage building information models (HBIM), multimodal data fusion, and digital twins. The evidence indicates a transition from isolated data-driven analysis toward predictive and decision-support systems, while persistent limitations remain in model generalization, benchmark datasets, semantic-to-HBIM transformation, multimodal fusion, and real-world deployment. Machine learning has substantial potential to support preventive, interpretable, and context-aware conservation, but future research requires more transparent reporting, shared datasets, external validation, and closer integration with conservation expertise.