Artificial intelligence for net-zero construction: a bibliometric mapping of research landscape, thematic clusters and emerging directions across the AEC industry

1 Introduction

The architecture, engineering, and construction (AEC) industry occupies a unique position in the global climate crisis. According to the International Energy Agency (IEA), buildings and construction together account for approximately 39% of global CO₂ emissions and consume roughly 40% of worldwide final energy (IEA, 2022). The Intergovernmental Panel on Climate Change (IPCC) has further documented that the building sector carries an estimated abatement potential of 8.2 GtCO₂e by 2050, one of the largest of any economic sector (Cabeza et al., 2022). Against this backdrop, achieving net-zero carbon targets across the full AEC value chain—from material production and design through construction, operation, and end-of-life—has become an urgent scientific and policy priority.

This value chain extends beyond individual buildings into urban-scale design: computational and generative design approaches have, for instance, been applied to optimise flood-resilient urban housing typologies in climate-vulnerable contexts (Maksoud et al., 2024), illustrating how sustainability-oriented computational methods operate across scales that a building-only review would not capture. Major economies including the European Union, the United Kingdom, China, and the United States have committed to carbon neutrality by 2050–2060, explicitly mandating radical transformation of the built environment sector (Sepasgozar et al., 2021).

Artificial intelligence (AI) has rapidly emerged as a transformative enabling technology for this transition. Machine learning (ML) and deep learning (DL) techniques have demonstrated significant potential for reducing energy consumption and carbon emissions across the AEC lifecycle. Applications range from predictive energy modelling and building performance simulation, to construction material optimization, embodied carbon quantification, and computer vision-based construction monitoring (Arsiwala et al., 2023; Gan et al., 2020).

Reinforcement learning algorithms have shown promise for dynamic energy control in smart buildings, while life cycle assessment (LCA) workflows are increasingly being automated through AI pipelines (Pinto et al., 2022). The aggregate potential of AI-driven interventions has been estimated at emission reductions of 8–40% depending on application domain, scale, and maturity of deployment (Thilakarathna et al., 2020).

The research landscape at the intersection of AI and net-zero construction has experienced an extraordinary expansion since 2020. Three structural forces converged: (i) a wave of national and institutional net-zero commitments following the COVID-19 recovery period; (ii) the rapid maturation of deep learning architectures and their extension to construction-relevant data types (e.g., Building Information Modelling (BIM) data, sensor streams, point clouds); and (iii) a significant increase in the availability of curated AEC-specific datasets enabling supervised learning at scale. More recently, the emergence of large language models (LLMs) and generative AI from 2023 onwards, exemplified by foundation models capable of automated design generation, LCA automation, and project management assistance, has opened a new frontier that existing reviews have not systematically captured (Ding et al., 2024).

Notwithstanding this growth, the bibliometric mapping of AI for net-zero construction remains limited and incomplete. The most directly comparable prior work is Li et al. (2024), published in Buildings, which constitutes the first systematic science mapping review of AI applications for net-zero carbon emissions. However, that study encompasses only 154 articles, restricts its scope to “sustainable building projects” (thereby excluding infrastructure, supply chains, project management, and end-of-life dimensions), relies exclusively on the Scopus database, and predates the 2023–2025 generative AI and LLM surge.

Other relevant reviews cover narrower sub-domains: AI for construction safety and quality control (Wang et al., 2024); ML for building energy efficiency (Lu et al., 2021); AI and structural analysis (Munoz-La et al., 2021). None provides an integrated view of the entire AEC industry oriented toward net-zero objectives, nor do any include the period of generative AI adoption (2023–2026).

A critical research gap therefore persists: among the reviews discussed above, none has produced a comprehensive bibliometric mapping of the triple intersection of AI + net-zero construction + full AEC industry, using dual databases (Scopus and Web of Science), covering the period 2015–2026, and explicitly analysing the emerging role of LLMs and generative AI. Given the pace at which this literature is expanding, closely related work may have appeared concurrently with the preparation of this manuscript. This gap limits researchers’ ability to identify the intellectual structure of the field, the most productive research communities, the dominant thematic clusters, and the emerging directions that should prioritise the next research agenda.

This study addresses this gap through a bibliometric and science mapping analysis, in which document identification and selection followed a systematic, PRISMA-informed protocol; the study does not undertake qualitative synthesis or methodological quality appraisal of individual included studies, which fall outside the scope of a bibliometric design. The following research questions (RQs) structure the investigation:

RQ1: What are the growth patterns and geographical and institutional distribution of research on AI for net-zero construction?

RQ2: What is the intellectual structure of the field in terms of most influential authors, journals, and documents?

RQ3: What thematic clusters define the current knowledge landscape of AI for net-zero AEC?

RQ4: What are the emerging research fronts (especially post-2022) and what persistent gaps require future attention?

The contributions of this article are threefold. First, relative to the reviews identified in Section 1, it provides a broader and more current bibliometric mapping of AI for net-zero construction, encompassing the full AEC industry and the generative AI era. Second, it applies a dual-database PRISMA protocol to a corpus more than three times larger than the most comparable prior review (Li et al., 2024); this difference reflects both the broader scope (full AEC industry rather than building projects alone) and the extended time window (through 2026) rather than broader inclusion criteria alone, as detailed in Section 4.2. Third, this synthesis appears in Frontiers in Sustainability, situating the AI-for-net-zero-construction intersection within a venue focused on sustainability outcomes and decision-making rather than computational methods as a primary discipline.

The remainder of this article is structured as follows: Section 2 describes the materials and methods; Section 3 presents the results across six analytical dimensions; Section 4 discusses the implications, limitations, and future directions; and Section 5 draws conclusions aligned with each RQ.

2 Materials and methods

Bibliometric analysis is a quantitative method for evaluating large bodies of scientific literature through the application of statistical techniques to bibliographic data (Donthu et al., 2021). Unlike a conventional systematic review, this design does not involve qualitative synthesis or critical appraisal of individual studies; PRISMA guidelines are applied here specifically to the document identification and selection stage, ensuring a transparent and reproducible corpus assembly process. It is particularly well-suited to emerging, rapidly expanding research fields where the volume of publications renders traditional narrative reviews insufficient (Aria and Cuccurullo, 2017).

The present study combines performance analysis (productivity and citation metrics) with science mapping (network analysis and thematic mapping) to characterise both the visible and structural dimensions of the AI–net-zero–AEC literature. This dual approach follows the methodological framework recommended by Aria and Cuccurullo (2017) and has been adopted in analogous bibliometric reviews across engineering and sustainability disciplines (Khahro et al., 2021). All analyses were performed with the bibliometrix R-package (v5.1.1) (Aria and Cuccurullo, 2017) using the biblioshiny interactive interface for visual exploration and export.

Data were retrieved from two complementary bibliographic databases: Scopus (Elsevier) and the Web of Science (WoS) Core Collection (Clarivate Analytics). The dual-database approach mitigates individual-database coverage gaps and increases corpus representativeness, particularly for interdisciplinary fields such as AI and sustainable construction (Aria and Cuccurullo, 2017). Searches were executed and data downloaded on 26 May 2026. Citation counts reported throughout this study (Section 3.6, Table 1) reflect Scopus- and WoS-indexed citations as recorded at this same retrieval date; as citation counts change continuously, figures reported here should not be assumed current beyond this date.

Table 1. Top 15 most globally cited documents in the corpus.

Global Citations retrieved on 26 May 2026 from Scopus and Web of Science; TC/Year computed as Global Citations divided by document age in years as of the retrieval date.

The primary search equation applied in Scopus was:

TITLE-ABS-KEY((“artificial intelligence” OR “machine learning” OR “deep learning” OR “neural network” OR “computer vision” OR “natural language processing” OR “large language model” OR “generative AI”) AND (“net-zero” OR “net zero” OR “carbon neutral” OR “zero carbon” OR “decarbonization” OR “carbon emission” OR “carbon footprint” OR “low carbon construction”) AND (“AEC industry” OR “architecture engineering construction” OR “construction industry” OR “built environment” OR “building sector” OR “construction project*” OR “sustainable construction” OR “green building”)) AND (LIMIT-TO(DOCTYPE, “ar”) OR LIMIT-TO(DOCTYPE, “re”)) AND (LIMIT-TO(LANGUAGE, “English”)) AND (PUBYEAR > 2014).

An equivalent structured query was deployed in

the Web of Science Core Collection using the TS = (Topic Search) field operator:

TS = ((“artificial intelligence” OR “machine learning” OR “deep learning” OR “neural network” OR “computer vision” OR “natural language processing” OR “large language model” OR “generative AI”) AND (“net-zero” OR “net zero” OR “carbon neutral” OR “zero carbon” OR “decarbonization” OR “carbon emissions” OR “carbon footprint” OR “low carbon construction”) AND (“construction industry” OR “built environment” OR “building sector” OR “construction project” OR “sustainable construction” OR “green building” OR “AEC industry” OR “architecture engineering construction”)) AND DT = (Article OR Review) AND LA = (English) AND PY = (2015–2026).

In both databases, the search was restricted to peer-reviewed Articles and Reviews published in English from 2015 onwards. Although the search protocol was executed across the full 2015–2026 window, no records satisfying the inclusion criteria were retrieved for 2015 or 2016; the earliest qualifying publication in the final corpus dates from 2017. This indicates that research explicitly intersecting AI, net-zero targets, and the AEC industry had not yet appeared in indexed literature during the initial two years of the search window, consistent with the Gestation-phase pattern described in Section 3.1.1.

Records exported from Scopus and WoS (in BibTeX/plain-text format) were imported and merged using the bibliometrix convert2df() and mergeDbSources() functions, which identify duplicate records across databases through matching of normalised DOI strings and, where DOI was unavailable, a combination of title and publication year similarity. This automated merge identified 253 duplicate records (Table 2), which were removed prior to screening.

Table 2. Search parameters and corpus assembly.

Title and abstract screening against the inclusion and exclusion criteria below was conducted independently by two of the authors; a third author adjudicated the small number of disagreements (fewer than 5% of screened records), following discussion until consensus was reached. Article identification and selection followed a PRISMA-inspired protocol for transparent, reproducible corpus assembly (Figure 1); as detailed below, records were subsequently assigned to two tracks, a quantitative corpus and a narrative-synthesis subset, described below.

Inclusion criteria required: (i) peer-reviewed articles or reviews indexed in Scopus or WoS; (ii) explicit intersection of AI/ML techniques with net-zero, decarbonisation, or carbon emission targets in the AEC industry or built environment; (iii) English language; and (iv) publication year 2015–2026. Document type, language, and publication-year criteria (conference proceedings, non-English records, and out-of-range publication years) were enforced directly through the database-level filters embedded in the search equations (Section 2, LIMIT-TO(DOCTYPE), LIMIT-TO(LANGUAGE, “English”), PUBYEAR > 2014), such that the 773 records retrieved from Scopus and WoS (Table 2) had already satisfied these criteria at export.

Consequently, the only screening step applied after retrieval was the removal of inter-database duplicates (n = 253), identified as described in Section 2, yielding the final corpus of 520 documents. No records were additionally excluded at the title/abstract or full-text screening stages, as the database-level filters and the topical specificity of the search equation (Section 2) had already restricted retrieval to records meeting the full inclusion criteria; abstract availability and topical relevance to AEC were verified as a quality check on the retrieved set rather than as a separate exclusionary screening pass.

Of the 520 records that passed deduplication, three were identified as retracted publications at the time of data retrieval (26 May 2026) (Journal of Intelligent & Fuzzy Systems (2022), Energy (2025), and Journal of Sensors (2022)) and were excluded from the analytical corpus. The final corpus therefore comprises 517 documents. In addition, a purposive subset of the most-cited documents (n = 15, Table 1, ranked solely by global citation count) was examined to identify documents suited to individual narrative synthesis of the field’s intellectual structure (Section 3.6). Of these fifteen, eleven were selected because each anchors a distinct segment of the AEC value chain, a distinct phase of the field’s temporal evolution, or serves as a scope-delimiting comparator reference (Section 1); the remaining four (Ashraf, 2022; Torabi, 2017; Liu, 2023; Gao, 2023) address a facet already exemplified by another profiled document and are reported quantitatively in Table 1 without repeated narrative treatment (Figure 1). This selection determines only the depth of narrative elaboration; it does not affect the citation ranking itself, nor any quantitative analysis reported in Section 3, both of which are based on the full n = 517 corpus or the full n = 15 anchor set as applicable. Table 2 summarises the search parameters.

The analyses conducted and the corresponding bibliometrix functions are summarised in Table 3. Performance analysis encompassed annual publication trends (biblioAnalysis()), Bradford’s Law for journal core identification (bradford()), Lotka’s Law for author productivity (lotka()), and citation analysis.

Table 3. Bibliometric analyses performed and corresponding bibliometrix functions.

Science mapping included keyword co-occurrence networks (networkPlot(type = “keyword”)), author collaboration networks (networkPlot(type = “collaboration”)), the Thematic Map using Callon’s centrality–density approach (thematicMap()), and thematic evolution across three sub-periods (2017–2020, 2021–2022, 2023–2026) using thematicEvolution(). Network community detection employed the Louvain algorithm. All network visualisations were generated and exported from biblioshiny; the co-occurrence network used a minimum keyword co-occurrence threshold of five.

Keyword co-occurrence analysis (Figure 2), thematic mapping (Figure 3), and thematic evolution analysis (Figure 4) were conducted on the DE field (author keywords), consistent with the manuscript’s distinction between author keywords and algorithmically-derived Keywords Plus (Section 3.4.1). Prior to network construction, a documented synonym-merging procedure consolidated author-keyword variants — including acronym/full-term pairs (e.g., ‘AI’/'artificial intelligence’), index-splitting artifacts (e.g., ‘machine’ and ‘learning’ appearing as separate entries from ‘machine learning’), and spelling variants (e.g., ‘building information modeling’/'modelling’) — using the synonyms parameter of bibliometrix’s biblioNetwork(), thematicMap(), and thematicEvolution() functions. At the applied minimum co-occurrence threshold of five, network modularity is comparatively low (Q ≈ 0.27–0.28), reflecting that only two densely connected communities emerge algorithmically alongside a long tail of loosely-connected peripheral terms (Section 3.4.2).

3 Results

3.1 Overview of scientific production

3.1.1 Annual publication trend

Of the 520 records that passed deduplication (Figure 1, Track 1), three were identified as retracted publications at the time of data retrieval (26 May 2026) (Journal of Intelligent & Fuzzy Systems (2022), Energy (2025), and Journal of Sensors (2022)) and were excluded from the analytical corpus. The final corpus therefore comprises 517 documents published between 2017 and 2026, encompassing 429 original articles (83.0%), 82 reviews (15.9%), four early-access articles, and one early-access review. Retraction status was verified against each record’s indexed status in Scopus and WoS as of the retrieval date; we note that retraction status is itself time-dependent, and future replications of this search may retrieve a different retraction count.

The annual growth rate (AGR), computed using the bibliometrix compound annual growth rate (CAGR) formula, stands at 54.45%. This figure is high in absolute terms, though direct comparison across bibliometric studies must account for differing corpus timespans, as AGR is sensitive to window length and tends to be inflated in shorter, more recent windows. Against this caveat, the 54.45% rate reported here is comparable to AGR values reported in other AI application domains undergoing active expansion, such as AI applied to electric vehicles (56.4%; Niu and Dobretsov, 2026) and AV trust and risk research (25.40–62.56% across sub-themes; Naiseh et al., 2025), and substantially exceeds the AGR typically reported for more established AI application domains such as AI in medicine (28.4% over a 2019–2023 window; Lin et al., 2025).

This positions AI for net-zero AEC among the faster-growing AI application fields currently under bibliometric study, without implying that its growth rate is categorically unprecedented. Figure 5 illustrates the temporal evolution of the corpus. Three distinct phases are evident. A Gestation phase (2017–2020) accounts for only 17 documents (3.3% of corpus), reflecting nascent interest in the intersection of AI and net-zero construction at a time when ML techniques were not yet mature enough for widespread AEC adoption and global net-zero commitments were still being formulated.

A Take-off phase (2021–2022) contributed 48 documents (9.3%), coinciding with the wave of national net-zero pledges following COP26 (Glasgow, 2021) and the accelerating adoption of deep learning in engineering applications. An Explosion phase (2023–2026) accounts for 452 documents (87.4% of the entire corpus), driven primarily by the convergence of urgent net-zero policy timelines and the mainstreaming of deep learning architectures for building and material performance prediction.

The parallel emergence of generative AI and LLMs from 2023 onwards contributes a fast-growing but, as Section 3.5 shows, still thematically peripheral share of this output. Notably, 2025 recorded the single largest annual output (n = 198), followed by 2026 (n = 100, likely underestimating the full-year total due to indexing lag at the time of retrieval). The average document age of 1.74 years and average citations per document of 17.67 confirm the corpus’ recency and its early but growing citation impact.

3.1.2 Core journals – Bradford’s Law

The 517 documents were published across 213 distinct sources, indicating a moderate level of dispersion. Bradford’s Law analysis identified eight journals as the Zone 1 core (Table 4), collectively publishing 172 articles (33.3% of the total corpus). Figure 6 shows the Bradford distribution. The core is dominated by MDPI journals (Buildings, n = 33; Sustainability, n = 21; Energies, n = 12), reflecting the open-access publishing strategies of Chinese-affiliated researchers who constitute the largest national contributor (see Section 3.3).

Table 4. Bradford’s Law Zone 1 core journals (n = 8, covering 33.3% of corpus).

Elsevier journals of high impact—Energy and Buildings (n = 30), Journal of Building Engineering (n = 29), Journal of Cleaner Production (n = 19), Building and Environment (n = 16), and Applied Energy (n = 12)—constitute the remaining core. Notably, Automation in Construction, the leading AEC-specific engineering journal, appears at rank 16 in Zone 2 (5 articles), suggesting that the field’s core literature is anchored more strongly in energy and sustainability journals than in process-oriented construction engineering outlets.

3.2 Author and institutional analysis

3.2.1 Most productive authors

The 517 documents involve 2,023 authors with an average of 4.63 co-authors per document, indicating a collaborative field. Twenty-two documents (4.3%) are single-authored. Table 5 presents the 15 most productive authors by total documents. Zhang X (a) leads with 11 documents (fractionalized count: 2.17), followed closely by Zhang S (10 documents, 3.08 fractionalized), both unaffected by the exclusion of retracted records. Table 5 explicitly labels the two ‘Zhang X’ entries as (a) and (b), corresponding to two bibliographically distinct author identities that could not be merged through surname-plus-initial matching alone.

Table 5. Top 15 most productive authors in the corpus.

This is a known limitation of automated author disambiguation in bibliometric software when applied to naming conventions common in Chinese scholarly output (Aria and Cuccurullo, 2017), where a small set of surnames and single-character given names is shared by a very large author population. We did not attempt manual disambiguation via full-name or affiliation-based reconciliation, as this was not feasible at the scale of the full corpus (2,023 authors); consequently, all author-level productivity figures in this study should be interpreted as indicative of the dataset’s automated name-matching rather than as verified individual counts. Fractionalized author counts, which distribute credit proportionally across co-authors, mitigate but do not eliminate this limitation. The top 15 authors are predominantly China-affiliated researchers, consistent with China’s dominant national output (Section 3.3).

Author productivity in the corpus follows Lotka’s Law of scientific productivity, with an estimated exponent β = 3.26 (R² = 0.973), higher than the theoretical value of 2.0 associated with the classical inverse-square distribution. This indicates that author productivity in this field is more concentrated among single-paper contributors than the classical Lotka pattern predicts: 88.2% of the corpus’s 2,023 authors (n = 1,785) published exactly one document, compared with 65.4% predicted under the theoretical β = 2.0 model, while the proportion of authors with two or more documents falls off correspondingly faster than the classical benchmark. This pattern is consistent with the fragmented, nascent-field topology already observed in the author collaboration network (Section 3.2.2).

3.2.2 Author collaboration network

The author co-authorship network (Figure 7) presents a fragmented topology characteristic of nascent interdisciplinary fields. Five discrete colour-coded clusters are discernible, with Wang S serving as the primary hub in the corrected corpus (betweenness centrality = 132.75), followed by Zhou M (130.75) and Huang J (123.75). All major clusters contain exclusively Chinese-affiliated authors, implying that while the field is internationally active (17.31% international co-authorship rate), the dense intra-national collaboration within China does not yet extend across institutional and national boundaries. The isolated components and chain-like structures in the lower clusters indicate emerging research groups that have not yet integrated into the broader network. This fragmentation represents both an opportunity (for establishing cross-national collaborations) and a risk, as knowledge exchange between clusters is currently limited.

3.3 Geographical analysis

3.3.1 Country scientific production

The geographical distribution of scientific output is highly concentrated. Table 6 presents the top 20 countries by total affiliation-counted publications. China dominates with 562 affiliation counts, more than four times the output of the United States (n = 116) and the United Kingdom (n = 106). India ranks fourth (n = 84), followed by Saudi Arabia (n = 59) and Australia (n = 58). The world map (Figure 8) visually underscores this concentration: the Asia-Pacific region, particularly China, accounts for a disproportionate share of global output, while sub-Saharan Africa, Latin America (with the exception of Ecuador, rank 20, n = 10), and Central Asia are largely absent from the research landscape—a North–South gap with significant implications for the applicability of AI-driven net-zero solutions in low- and middle-income countries.

Table 6. Top 20 countries by total affiliation-counted publications.

3.3.2 Corresponding author countries and collaboration patterns

Analysis of corresponding author countries (Figure 9; Table 7) reveals that China contributes 199 corresponding-author articles (38.5% of corpus), confirming its role as the field’s primary knowledge-generating hub. India (7.3%), USA (6.9%), and Australia (5.4%) follow. The Single-Country Publication (SCP) vs. Multiple-Country Publication (MCP) ratio provides an indicator of international collaboration intensity. Egypt exhibits the highest MCP rate (50.0%), followed by Saudi Arabia (47.1%), suggesting that these countries preferentially engage in international collaborations rather than publishing domestically. In contrast, China’s MCP rate is 15.6% (SCP = 168, MCP = 31), USA 8.3%, South Korea and Germany 0%, indicating highly self-contained national research communities. The Country Collaboration Map (Figure 10) confirms that China serves as the gravitational node of the international collaboration network, with the strongest ties directed toward the USA, Australia, the UK, and selected Asian partners.

Table 7. Corresponding author countries: SCP/MCP collaboration patterns (top 12).

3.4 Keyword co-occurrence analysis

3.4.1 Most frequent keywords

The author keyword vocabulary comprises 1,813 unique terms (DE field), complemented by 2,447 Keywords Plus terms (ID field) algorithmically derived from reference titles. The five most frequent author keywords are: machine learning (n = 173), carbon emissions (n = 72), construction industry (n = 70), artificial intelligence (n = 69), and compressive strength (n = 63). The prominence of “compressive strength” (a concrete and materials science term) in the top five signals that a significant sub-corpus focuses on ML-based prediction of construction material properties as a pathway to reducing embodied carbon. This finding contrasts with the original expectation that operational energy efficiency terms would dominate and is discussed further in Section 4.

3.4.2 Co-occurrence network and thematic clusters

At the applied co-occurrence threshold (minimum frequency = 5), the keyword co-occurrence network (Figure 2) resolves into two substantively connected clusters rather than a larger number of comparably weighted ones, alongside a long tail of loosely-connected peripheral terms — a structure consistent with the network’s comparatively low modularity (Q ≈ 0.27–0.28) at this threshold.

Cluster 1 – Materials, Optimisation, and Carbon Quantification (red, 16 terms): Centred on machine learning as the dominant hub, this cluster encompasses carbon emission, artificial neural network, compressive strength, built environment, multi-objective optimization, sustainable construction, construction industry, optimization, embodied carbon, energy consumption, geopolymer concrete, carbon footprint, prediction, recycled aggregate concrete, and durability. It is the largest and most internally connected grouping in the network, reflecting the volume of research applying ML jointly to material performance prediction, carbon accounting, and optimisation-based design.

Cluster 2 – AI, Sustainability, and Digital Integration (blue, 6 terms): Grouped around artificial intelligence, this cluster connects sustainability, digital twin, energy efficiency, building, and building information modeling, reflecting the integration of AI with broader digital-transformation and sustainability framing in the AEC industry.

The remaining terms in the co-occurrence network, including deep learning, decarbonization, green building, building energy, environmental sustainability, machine learning algorithms, life cycle assessment, circular economy, smart buildings, digital twins, and climate change, do not meet the minimum co-occurrence threshold for cluster membership and appear as isolated or weakly-connected nodes. This does not indicate these themes are absent from the literature (several appear with substantial individual frequency; Section 3.4.1), but that they have not yet consolidated into densely interlinked keyword groupings at the author-keyword level, consistent with a field still in the process of thematic crystallisation.

3.4.3 Thematic evolution across three periods

The thematic evolution analysis (Figure 4) tracks the migration of keywords across three analytical windows: 2017–2020 (Gestation), 2021–2022 (Take-off), and 2023–2026 (Explosion), documenting a clear three-phase paradigmatic trajectory.

2017–2020 (Gestation): Author-keyword co-occurrence in this period is limited to four terms — artificial intelligence, artificial neural network, built environment, and climate change — reflecting early-stage scoping and conceptual positioning rather than applied implementation, consistent with the low document volume in this phase (Section 3.1.1).

2021–2022 (Take-off): Machine learning becomes the dominant term (relative weight 0.46), connecting to artificial neural network, carbon footprint, climate change, compressive strength, energy, and environmental impact. This marks the transition from conceptual framing toward applied prediction and materials-oriented research, though the still-limited term set (seven terms) indicates the field’s vocabulary had not yet diversified substantially.

2023–2026 (Explosion): The thematic core consolidates around machine learning and artificial intelligence (relative weights 0.38 and 0.31, respectively), connected to carbon emission, energy, building energy, and built environment. This narrower, more concentrated core — compared with the broader vocabulary hypothesized in earlier drafts of this analysis — indicates that, at the author-keyword level and after synonym consolidation, the field’s thematic evolution has proceeded through progressive concentration around a small set of foundational AI and carbon-related terms, rather than diversification into the wide range of specialised sub-themes (e.g., Bayesian optimisation, digital twins, interpretable ML) that remain present in the corpus but have not yet reached the co-occurrence threshold required to register as active nodes in this three-period comparison.

3.5 Thematic map (Callon’s centrality–density framework)

The Thematic Map (Figure 3) positions research clusters according to their Relevance degree (centrality, x-axis) and Development degree (density, y-axis), yielding four quadrants with distinct strategic interpretations (Table 8).

Table 8. Thematic map quadrant interpretation for the AI–net-zero–AEC field.

3.6 Most cited documents and intellectual structure

Table 1 presents the 15 most globally cited documents in the corpus, ranked solely by global citation count. Following the selection criterion set out in Section 2 and illustrated in Figure 1 (Track 2), eleven of these fifteen receive individual narrative discussion because each anchors a distinct AEC value-chain segment, temporal phase, or scope-delimiting comparator role not already exemplified by another profiled document; the remaining four (Ashraf, 2022; Torabi, 2017; Liu, 2023; Gao, 2023) address a facet already covered by a profiled document and are reported here in tabular form only. Among the eleven, the intellectual structure of the field is anchored by five particularly seminal contributions.

Yu et al. (2021), published in IEEE Trans Smart Grid with 413 citations (68.8 citations/year, the highest citation velocity in Table 1), represents the most cited document, applying IoT-enabled AI frameworks to smart building energy management. Its citation lead over more recent documents should be interpreted with some caution: as a 2021 publication, it has had five years to accumulate citations, compared to one to three years for the generative-AI-related documents discussed in Section 3.5, so part of this gap reflects citation lag rather than a difference in underlying influence. That said, its citation velocity remains among the highest in the corpus even when age-adjusted, suggesting that the AI + IoT + energy efficiency nexus it established has sustained rather than merely accumulated relevance. Sepasgozar et al. (2021) (254 citations, Buildings) provides a systematic review of digital twin applications in smart and sustainable buildings, anchoring the digital twin cluster.

Pinto et al. (2022) (226 citations, Advances in Applied Energy) contributes a deep reinforcement learning framework for building energy control, directly addressing net-zero operational performance. Wu et al. (2019) (172 citations, Transportation Research Part D) extends the field beyond buildings into transportation infrastructure carbon prediction, demonstrating the AEC-wide scope of the research problem. Arsiwala et al. (2023) (141 citations, Energy and Buildings), the most cited post-2022 document, applies machine learning methods to predictive monitoring of CO₂-equivalent emissions from existing buildings, representing the emerging paradigm synthesis. The remaining documents in the top 15 span LCA and embodied carbon (Gan et al., 2020; Thilakarathna et al., 2020), ML for concrete compressive strength prediction (Mahjoubi et al., 2021), and AI-driven waste management in construction (Lu et al., 2021).

4 Discussion

4.1 Growth trajectory and geographical distribution

The 54.45% annual growth rate documented in this corpus is high relative to comparable bibliometric studies and exceeds the growth rates reported in adjacent fields such as AI for construction safety (approximately 30–40% AGR) and BIM adoption research (Wang et al., 2024). The three-phase trajectory (Gestation, Take-off, Explosion) mirrors patterns observed in other transformative technology-meets-policy intersections, where sustained exponential growth is triggered by external catalysts. In this case, the COP26 net-zero pledges (2021) and the ChatGPT/LLM emergence (late 2022) function as sequential catalysts, explaining the two inflection points visible in Figure 5.

China’s dominance (38.5% of corresponding-author documents) warrants a more critical reading than volume alone provides. Part of this concentration plausibly reflects genuine research leadership: China’s dual carbon strategy (carbon peak by 2030, carbon neutrality by 2060) has translated into sustained, centrally coordinated funding for AI and green construction research, and Chinese research output in AI-related AEC bibliometric studies has been independently documented in comparable reviews (Li et al., 2024), suggesting the pattern is not an artefact of database coverage.

However, a substantial share of this concentration is also structural rather than purely scientific: China’s construction sector is, by a wide margin, the largest in the world by built floor area and annual output, which mechanically generates a correspondingly large volume of applied research regardless of per-capita research intensity or qualitative leadership. Distinguishing these two explanations is not possible from productivity counts alone and would require normalising output against construction-sector size or R&D investment, which lies beyond the scope of this bibliometric analysis; we flag this as an important caveat rather than a resolved finding.

The low MCP rate (15.6%) is itself analytically informative beyond simply indicating domestic orientation: it is consistent with a research ecosystem where funding, data access (e.g., proprietary BIM and sensor datasets), and publication incentives are structured nationally, and where language and institutional familiarity lower the transaction cost of within-country collaboration relative to cross-border partnerships.

The same structural logic plausibly explains the broader fragmentation observed in the author collaboration network (Section 3.2.2): the network’s five discrete clusters are not merely a statistical artefact of the Louvain algorithm but likely reflect genuine boundaries in funding streams, institutional review processes, and data-sharing regimes that differ by country. Left unaddressed, this fragmentation risks producing parallel, non-interoperable bodies of AI methodology for net-zero AEC—for instance, models trained and validated predominantly on Chinese building stock and climate data may embed assumptions that do not transfer cleanly to other construction contexts, a concern we return to in Section 4.4.

The contrasting pattern of Egypt (50% MCP) and Saudi Arabia (47.1% MCP) suggests that research from the Global South in this domain is systematically collaborative—these countries leverage international partnerships to compensate for domestic capacity limitations, a pattern with positive implications for research quality but which may also reflect dependency dynamics.

The near-absence of Latin America (10 articles from Ecuador constituting the only notable regional entry at rank 20), sub-Saharan Africa, and South and Southeast Asia beyond India is a critical finding. These regions face the largest infrastructure construction demands over the coming decades and are disproportionately vulnerable to climate change impacts.

The lack of domestic AI-for-net-zero-AEC research capacity risks importing solutions calibrated for high-income country construction contexts—a known limitation in technology transfer for sustainable construction (Sepasgozar et al., 2021). This North–South gap is a persistent structural problem that future research agendas must explicitly address.

4.2 Knowledge structure versus prior reviews

Comparison with Li et al. (2024) (the closest comparable study) reveals both continuities and critical divergences. It is important to clarify the source of the corpus-size difference (517 vs. 154 documents, approximately 3.4 times larger): approximately half of this difference is attributable to the broader scope adopted here (the full AEC industry rather than building projects alone, which alone accounts for the inclusion of Cluster 1, ML for Construction Materials, discussed below), and the remainder reflects the extended search window (through 2026, capturing the 2023–2026 Explosion phase that postdates Li et al.’s retrieval date) combined with dual-database coverage.

The difference therefore reflects genuine expansion of the underlying literature and of the review’s scope, rather than looser inclusion criteria; both reviews applied comparable PRISMA-based screening standards. The six mainstream topics identified by Li et al.—LCA and carbon footprint; practical AI applications; multi-objective optimisation; energy management and efficiency; carbon emissions from buildings; and decision support systems—map broadly onto Cluster 1 (materials, optimisation, and carbon quantification) of the present co-occurrence network, with energy management and decision-support themes further represented among the peripheral, sub-threshold terms discussed above.

However, the present study identifies an additional cluster, Materials, Optimisation, and Carbon Quantification (Cluster 1), that does not appear in Li et al. because material performance prediction, while central to embodied carbon reduction across the AEC supply chain, lies outside the “building project” scope adopted by that review. This cluster is the largest in the network by node count and internal connectivity, underscoring the limitation of restricting bibliometric scope to buildings alone.

The Thematic Map (Figure 3) further reveals dynamics invisible to Li et al.: the appearance of generative artificial intelligence in the Niche quadrant (centrality = 0.00, density = 62.5, both computed relative to the corpus median) signals that this technology stream is internally accumulating mass while not yet being broadly cross-referenced with the field’s established themes. This is precisely the early-stage signature of a paradigm shift. Explainable AI and digital twin terms, in contrast, do not form independent clusters at the applied co-occurrence threshold; they are absorbed into the dominant ‘artificial intelligence’ cluster (85 terms, centrality = 8.76, the highest in the network), which falls in the Basic quadrant — indicating that interpretability and digital-twin concerns are discussed as part of the field’s broad, high-relevance AI vocabulary rather than as a distinct, self-standing thematic front.

Neither of these dynamics could have been captured in a study with a corpus retrieval date predating 2023. It should be noted that generative AI’s presence in the Niche rather than Motor quadrant indicates that, despite its growing publication volume, it has not yet been broadly cross-referenced with the field’s established themes; its contribution to the 2023–2026 growth documented in Section 3.1.1 is therefore one of volume, not yet of structural centrality.

The thematic evolution analysis provides the most compelling evidence for the paradigmatic transition the field has undergone. Between 2017–2020 and 2023–2026, the primary knowledge currency shifted from classical ML for building energy performance to a multi-dimensional knowledge ecosystem encompassing net-zero carbon as an explicit optimisation target, Bayesian and stochastic methods for uncertainty quantification, with digital twin integration visible in the co-occurrence network’s Cluster 2 (Section 3.4.2) though not yet a high-frequency term in the period-by-period keyword evolution, and the emergence of LLM-compatible workflows. This trajectory is consistent with broader technology forecasts for the role of foundation models in decarbonisation (Ding et al., 2024) but had not been documented bibliometrically for the AEC domain.

A related methodological caveat concerns the interpretation of Table 1 itself. Citation counts inherently favour earlier publications: Yu et al.'s (2021) 413 citations reflect five years of accumulation, whereas the most-cited document from the 2023–2026 Explosion phase, Arsiwala et al. (2023), has accrued 141 citations in roughly three years (35.3 citations/year).

When ranked by citation velocity rather than raw count, several recent documents approach or exceed Yu et al.’s rate, indicating that citation momentum for net-zero-specific and digital-twin-integrated AI applications is building quickly even though raw counts still favour earlier IoT-centric work. This distinction matters for readers using Table 1 to gauge which lines of enquiry are currently most active, as opposed to which have been historically most cited.

4.3 Theoretical implications

The co-occurrence network and Thematic Map together are consistent with a field beginning a transition from narrow AI (task-specific, domain-constrained ML models) toward applied general AI (multi-modal, generalisable, LLM-enabled systems), although this characterisation should be treated as an early-stage signal rather than an established trajectory, given that the corpus captures only the first three to four years of this shift and generative AI itself remains in the Niche quadrant of the Thematic Map (Section 3.5).

The convergence of BIM, digital twins, IoT, and AI (visible in Cluster 2’s grouping of digital twin, building information modeling, and artificial intelligence (Section 3.4.2), though not yet reflected as an active node in the three-period thematic evolution comparison (Section 3.4.3)) is suggestive of an emerging methodological paradigm that could, if the trend continues, reorganise the traditional AEC project lifecycle around continuous data-driven optimisation rather than discrete decision points. Should this trajectory consolidate in future literature, it would not be merely additive (more AI tools) but would restructure the epistemology of sustainable construction: performance increasingly defined and measured in real time, targets dynamically updated, and optimisation occurring across scales (material–component–building–district–city) simultaneously. We present this as a plausible interpretation of current evidence rather than a settled conclusion, and return to this point in Section 4.6.

The four vectors of AI’s impact on construction carbon emissions (design optimisation, construction process monitoring, operational energy control, and end-of-life/LCA) are unevenly represented in the corpus. Operational energy and material performance, both represented within Cluster 1, dominate; construction process monitoring and end-of-life applications remain underrepresented, as confirmed by the absence of computer-vision-for-site-monitoring and demolition-waste-AI keywords in the high-centrality quadrant of the Thematic Map. This imbalance suggests that the field’s most impactful future contributions may lie precisely in these underserved lifecycle phases.

4.4 Practical implications

Consistent with this section’s focus on modeling and optimization for decision support, the Thematic Map offers industry practitioners a quantitative, evidence-based prioritisation framework rather than a qualitative literature summary: Motor-quadrant themes (monte carlo simulation for carbon emissions in the building sector, ML-based energy consumption prediction) represent mature technologies with demonstrated centrality and density in the published record, and are therefore lower-risk candidates for near-term decision-support tool development, while the Niche quadrant (including generative AI and building energy retrofit) marks the high-value frontier for near-term R&D investment.

The fragmentation of the author collaboration network (Figure 7) has practical consequences: knowledge silos between Chinese, Western, and emerging-economy research communities limit the cross-pollination needed to adapt AI solutions to diverse construction contexts. Bridging these silos is not purely a matter of establishing formal research partnerships; it also depends on translational and pedagogical frameworks that connect academic findings with practice, such as experiential and site-based teaching approaches that expose future architects and engineers directly to context-specific design challenges (Yahia et al., 2023). Industry bodies (e.g., CIOB, AGC, FIDIC) and international research consortia should actively facilitate researcher exchange and joint dataset development across these clusters.

Beyond collaboration silos, practical deployment of AI-for-net-zero tools faces technical barriers largely absent from the bibliometric record: data privacy and access restrictions on proprietary BIM models and building-performance datasets, vendor lock-in around proprietary BIM schemas that impede cross-platform model exchange, and incomplete interoperability even where open standards such as IFC (Industry Foundation Classes) exist, since AI pipelines frequently require data transformations not natively supported by current IFC implementations. These barriers are not captured by co-occurrence or thematic-map analysis of published keywords, since they concern implementation practice rather than research topic, but they represent a material gap between the field’s academic maturity and its industrial deployment readiness.

For policymakers in emerging economies, the North–South gap documented in Section 3.3 is actionable evidence that capacity-building initiatives are needed. AI tools for net-zero AEC trained predominantly on Chinese or European building stock may perform poorly when applied to tropical, arid, or informal construction contexts. Investment in geographically adapted research programmes (particularly for India, sub-Saharan Africa, and Latin America) is not merely an equity concern but a technical necessity for achieving global net-zero targets across the full built environment.

4.5 Limitations

Several limitations of this study warrant acknowledgement, spanning both the scope of the corpus and the inherent constraints of bibliometric methods.

Regarding corpus scope: first, the corpus is restricted to peer-reviewed articles and reviews indexed in Scopus and WoS, excluding grey literature, conference proceedings, preprints, and patents. Innovative AI applications for net-zero construction are often first disseminated through conference papers (e.g., ISARC, ICCREM) or technical reports, meaning that industry-facing innovation may be underrepresented; while we cannot quantify the precise magnitude of this gap without a parallel search of conference databases, the concentration of core journals in energy and sustainability outlets rather than construction-process venues (Section 3.1.2) is one indirect signal that process- and site-level innovation may be underrepresented relative to material- and energy-focused research.

Second, the English-language restriction excludes substantial bodies of Chinese-language and other non-English publications, potentially underestimating outputs from non-English-speaking research communities even within the indexed literature — a limitation of particular relevance given China’s dominant position in the corpus (Section 3.3).

Third, citation-based indicators are subject to citation lag, whereby recently published articles (particularly from 2025 to 2026) have had insufficient time to accumulate citations regardless of their eventual influence; this likely affects the ranking of most-cited documents in Table 1 and should be considered when interpreting the apparent dominance of earlier (2019–2022) publications.

Fourth, author-level productivity metrics are affected by name-disambiguation limitations, as discussed in Section 3.2.1, which are especially pronounced for common Chinese surnames and may both overstate the productivity of some author identities and fragment the true output of others across multiple unmerged entries.

4.6 Future research directions

Because the generative AI and LLM trend documented here spans a short and still-unfolding publication window, the following directions are proposed with the explicit understanding that the field’s trajectory may evolve substantially before these questions can be conclusively answered; periodic bibliometric updates will be necessary to confirm whether the patterns identified in Section 3.5 and 4.3 consolidate, plateau, or reverse. With this caveat, the findings of this study suggest several high-priority directions for future research:

• (i) Patent analysis of AI applications for net-zero AEC would capture innovation occurring at the technology-commercialisation interface that peer-reviewed literature does not reflect; a concrete starting point would be to ask which AI-for-net-zero-AEC patent families have no corresponding peer-reviewed publication, and whether commercial patenting activity is geographically distributed differently from the publication patterns documented in Section 3.3.
• (ii) Regional bibliometric sub-analyses focusing on the Global South (particularly sub-Saharan Africa, Southeast Asia, and Latin America) are needed to map the localisation gap in AI-for-net-zero research, specifically addressing whether AI models validated on data from these regions perform comparably to those trained on Chinese or European building stock, and what data infrastructure gaps currently prevent such validation.
• (iii) The emergence of generative AI and LLMs in the Niche quadrant warrants a dedicated systematic review examining what specific tasks (BIM-based design generation, automated LCA, regulatory compliance checking) these tools have been applied to, with what performance benchmarks relative to non-generative baselines, and at what stage of technology readiness (proof-of-concept versus deployed pilot versus operational use).
• (vi) A meta-analysis comparing the empirical effectiveness of different AI techniques (ML, DL, reinforcement learning, generative AI) for reducing measured carbon emissions across building types and lifecycle phases should ask, specifically, which technique class delivers the largest measured emission reduction per unit of implementation cost, and whether this ranking is stable across climate zones and building typologies.
• (v) Mixed-method studies combining bibliometric mapping with expert interviews could validate the emerging research fronts identified here against practitioner and policymaker priorities, for example by testing whether industry practitioners in the AEC sector independently rank generative AI and explainable AI as high-priority investment areas, as the Thematic Map (Section 3.5) suggests, or whether practitioner priorities diverge from the publication-based signal.

5 Conclusion

This study presents a comprehensive bibliometric mapping of research on artificial intelligence for net-zero construction across the AEC industry, based on 517 peer-reviewed documents from Scopus and Web of Science (2017–2026, retrieved 26 May 2026, following exclusion of three retracted records). The findings provide evidence-based answers to the four research questions that structured the investigation.

Regarding RQ1 (growth and geography), the corpus exhibits a high annual growth rate of 54.45%, with 87.4% of all publications concentrated in 2023–2026, documenting a substantial research explosion driven by the convergence of net-zero policy commitments and the emergence of generative AI. China dominates output (38.5% of corresponding-author documents), while a pronounced North–South gap leaves the Global South—the region facing the largest future construction demand—largely absent from the research landscape.

Regarding RQ2 (intellectual structure), eight journals constitute the Bradford Zone 1 core (33.3% of corpus), dominated by MDPI and Elsevier sustainability-oriented outlets rather than construction-engineering journals. The most cited document (Yu et al., 2021, 413 citations) establishes AI+IoT + energy management as the field’s foundational intellectual anchor. The author network is fragmented across Chinese-dominated clusters with limited inter-national linkage, signalling a field that has grown rapidly within national silos rather than through globally integrated collaboration.

Regarding RQ3 (thematic clusters), keyword co-occurrence analysis identified two densely connected thematic clusters — materials, optimisation, and carbon quantification; and AI, sustainability, and digital integration — alongside a substantial set of peripheral, sub-threshold themes (including deep learning, decarbonisation, circular economy, and digital twins) not yet consolidated into distinct keyword groupings. The Thematic Map positions large language models, simulation, and interpretability methods as Motor themes (the productive core), while generative AI occupies the Niche quadrant, internally accumulating but not yet diffused across the field.

Regarding RQ4 (emerging fronts and gaps), the thematic evolution documents a clear three-phase paradigmatic shift from classical ML through deep learning + IoT to net-zero-specific AI with Bayesian and stochastic methods. Generative AI, LLMs, digital twins for real-time carbon monitoring, and interpretable ML represent the leading research frontier. Critical gaps persist in construction process monitoring, end-of-life AI applications, Global South adaptation, and empirical efficacy benchmarking.

The irruption of generative AI and LLMs in 2023–2026, while still occupying a Niche position in the field’s thematic structure, signals the early stage of a reorientation from prediction toward the generation and autonomous optimisation of net-zero construction solutions—an agenda that prior reviews, including Li et al. (2024), could not have anticipated and that future bibliometric updates should continue to track as this theme moves toward the field’s core. This study establishes an early bibliometric baseline for tracking this transformation and provides a reference for researchers, industry actors, and policymakers navigating the AI-enabled decarbonisation of the built environment.

For researchers, the Thematic Map and thematic evolution results (Sections 3.5–3.6) offer a concrete prioritisation tool: Motor-quadrant themes represent areas where incremental contributions are more likely to be rapidly cited and integrated into an active research conversation, while Niche-quadrant themes such as generative AI represent higher-risk, higher-novelty opportunities for researchers seeking to establish new research lines ahead of the field’s broader uptake.

For industry practitioners, the Bradford core journals (Table 4) and most-cited documents (Table 1) identify where mature, field-tested AI methodologies for net-zero AEC are most likely to be documented in sufficient technical detail for direct adaptation. For policymakers, the geographical concentration and North–South gap documented in Section 3.3 indicate that national AI-for-construction strategies calibrated on evidence from China, the USA, or the UK may require independent validation before being adopted as policy templates in construction contexts with different climates, material availabilities, and regulatory environments.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. Generative artificial intelligence tools were used during manuscript preparation to assist with language editing, formatting, and organization of the analysis narrative. All bibliometric data, statistical outputs, and interpretive conclusions were generated and verified by the authors using bibliometrix in R.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Chang He, Sun Yat-sen University, China

Reviewed by: Salem Baarimah, Hadramout University of Science and Technology, Yemen

Muhamad Abduh, Bandung Institute of Technology, Indonesia

Amany Salman, Prince Sultan University, Saudi Arabia