AI research on historic buildings and monuments has surged since 2020, review finds

Artificial intelligence has moved from the margins of heritage science to one of its fastest-growing research frontiers, according to a bibliometric review published in the journal Built Heritage. The study, led by Yangting Wen and Marzudi Md Yunus, analysed 382 English-language publications retrieved from the Web of Science Core Collection covering the period from 1 January 2015 to 1 October 2025. Its focus was deliberately narrow: immovable cultural heritage, meaning historic buildings, archaeological sites, monuments, cultural landscapes and other heritage places with fixed spatial locations. Movable objects, museum collections and intangible heritage were excluded unless the studies directly addressed the documentation, monitoring, conservation or management of an immovable site.

The scale of the recent boom is striking. Between 2015 and 2019, the annual number of publications in the dataset remained in the single digits, a period the authors describe as an early stage dominated by conceptual discussions and exploratory digital applications. After 2020 the curve bends sharply upward: 17 publications appeared in 2021, 33 in 2022 and 39 in 2023, before the count jumped to 103 articles in 2024. By 1 October 2025, 173 publications had already been retrieved, although the authors caution that this partial-year figure cannot be compared directly with complete annual totals. The cumulative trend follows a power-growth pattern, described by the equation y = 1.0053x^1.9927 with an R² of 0.7936, indicating sustained expansion within the captured literature.

Reaching that dataset required an unusually careful screening pipeline. The initial database search returned 562 records, which were reduced to 547 after restricting the results to English-language publications. Filtering by document type, retaining only articles and review articles, left 426 records for manual assessment. The authors then reviewed titles, abstracts, keywords and, where necessary, full texts, excluding 44 records for reasons including research objects outside the scope of immovable heritage, no substantive application of artificial intelligence, exclusive focus on movable collections or intangible heritage, and duplicate or incomplete bibliographic records. The final 382 publications were analysed using VOSviewer 1.6.19 for collaboration and keyword networks, CiteSpace 6.2.R6 for co-citation and burst detection, and ArcGIS 10.8.1 for geospatial mapping.

The geographical picture is heavily concentrated. China was the most productive country by a wide margin, contributing 168 publications, or 44 percent of the total. Italy followed with 65 publications (17 percent) and Spain with 31 (8 percent). The United States contributed 15 publications, while India, Malaysia, Portugal, South Korea and the United Kingdom each contributed 14. Countries in Africa and South America were rarely represented, a pattern the authors attribute partly to publication language, database coverage, research infrastructure and regional publication practices rather than to differences in research capacity alone. Because the analysis is based primarily on publication volume, the findings indicate research productivity rather than citation-adjusted academic influence.

Collaboration networks mirror this concentration. China maintains co-authorship links with Malaysia, France, Portugal, Australia, Greece, the United States, the United Kingdom, India and Türkiye, while Italy is connected with partners including the Netherlands, Sweden, Austria and Israel, and Spain collaborates with countries in South America and Asia, including Colombia, Chile, South Korea, India and Vietnam. The authors stress that these network positions reflect the extent and structure of collaboration within the dataset, not scholarly quality or leadership, and that some countries occupy peripheral positions with few international links for reasons that may include language and database visibility.

The disciplinary profile is dominated by technical fields. Materials Science, Multidisciplinary ranks first among Web of Science categories with 100 publications, followed by Chemistry, Analytical with 86 and Spectroscopy with 80, reflecting extensive use of AI in material characterisation, pigment and elemental analysis, deterioration detection and structural diagnosis. Computer science, geosciences, remote sensing, engineering and environmental science are also strongly represented. Archaeology, with 69 publications, Humanities, Multidisciplinary with 67 and Art with 60 also appear among the leading fields, showing that AI is applied to archaeological interpretation, stylistic analysis, digital reconstruction and archival processing. The journal landscape is similarly skewed: the Journal of Cultural Heritage leads with 58 publications and 186 citations, followed by npj Heritage Science with 21 publications, while Remote Sensing, Scientific Reports and other science-oriented venues feature prominently.

Co-citation analysis reveals the intellectual foundations of the field. The citation network generated with CiteSpace comprised 420 nodes and 975 citation links, with a network density of 0.0111. Two references stand out as the most influential nodes: Fiorucci and colleagues’ 2020 overview of machine learning in cultural heritage, with 199 citations, and Mishra and colleagues’ 2021 review of machine learning for the structural health monitoring of heritage buildings, with 177 citations. The latter covers applications ranging from predicting masonry and mortar properties to automatic damage detection and seismic vulnerability assessment, while also flagging persistent challenges such as limited data availability, weak model generalisation and sensor precision constraints.

Keyword mapping identifies five major thematic clusters that together define the field’s knowledge structure. The first centres on digital cultural heritage and immersive technologies, including virtual reality, augmented reality, digital twins and museums. The second addresses sustainability, structural monitoring and environmental risk, encompassing crack detection, climate-related impact assessment and long-term asset management. The third covers cultural heritage conservation, deterioration, digitisation, management and preservation. The fourth and fifth clusters focus on computational methods themselves, including machine learning, deep learning, convolutional neural networks, object detection, image recognition, photogrammetry, point-cloud analysis and semantic segmentation. Temporal keyword analysis shows the themes shifting from heritage preservation, preventive conservation and sensor-based monitoring before 2023, towards machine learning, computer vision and pattern recognition in 2023 and 2024, and then towards deep learning, digital twins, damage detection and metaverse-related approaches in 2024 and 2025.

Burst detection using Kleinberg’s algorithm on the top 30 keywords confirms this trajectory. Between 2020 and 2022, attention focused on heritage digitisation and 3D modelling, with digital heritage showing a burst strength of 3.1. From 2023, terms such as artificial intelligence, neural networks, virtual reality and models surged, alongside scan to BIM and non-destructive testing, indicating AI’s shift from a peripheral tool to a core methodological framework. From 2024 onward, strong bursts in classification, augmented reality, generative adversarial networks and stable diffusion suggest that deep learning and content generation have become the research frontier. The authors synthesise this evolution into three stages: digital documentation and data acquisition from 2015 to 2019, rapid integration of machine learning and computer vision from 2020 to 2022, and deeper integration with governance, sustainability frameworks and public engagement from 2023 to the present.

The review is equally notable for its caveats. The authors repeatedly warn that the concentration of publications in technical fields should not be read as evidence of a general disciplinary shift across cultural heritage studies, since the dataset reflects the technological orientation of an English-language Web of Science search strategy. Publication counts and network positions, they note, cannot rank countries by scholarly influence, and methodological sophistication in itself does not demonstrate improved heritage outcomes. They also highlight risks that AI models trained on geographically limited datasets may transfer poorly across architectural traditions and climates, potentially marginalising local knowledge systems, a concern they link to what has been described as algorithmic colonialism. Generative tools such as GANs and diffusion models raise further questions about authenticity and interpretive authority in reconstruction. The authors call for geographically diverse datasets, explainable AI, cross-regional validation and collaboration with heritage professionals and communities, and they note that future work should incorporate additional databases and multilingual publications to overcome the limitations of the present analysis.

Subject of Research: Bibliometric analysis of artificial intelligence applications in immovable cultural heritage research

Article Title: Artificial intelligence in immovable cultural heritage research: a bibliometric and knowledge mapping review

Article References: Wen, Y., & Yunus, M. M. (2026). Artificial intelligence in immovable cultural heritage research: a bibliometric and knowledge mapping review. Built Heritage, 10(1), Article 75. https://doi.org/10.1186/s43238-026-00324-1

Image Credits: AI Generated

DOI: 10.1186/s43238-026-00324-1

Keywords: artificial intelligence, cultural heritage, bibliometrics, built heritage, structural health monitoring, digital twins, deep learning, remote sensing, Web of Science, China, Italy, conservation science