Integrating Artificial Intelligence and BIM in Construction: Systematic Review and Quantitative Comparative Analysis

Abstract

In the transition toward a more digital and data-driven construction industry, understanding how Artificial Intelligence (AI) and Building Information Modeling (BIM) are integrated is key to planning, delivering, and operating projects effectively. This review examines recent studies to identify usage patterns of AI and BIM. Searches were conducted on the Web of Science Core Collection from 2022 to 2025. After running a reproducible review protocol aligned with PRISMA 2020, which began with 1212 articles, and after a funneling process, 12 studies met the predefined eligibility criteria. In the present study, the synthesis was non-meta-analytic; instead, the information was analyzed by using standardized tabulation with a consistent format and compared using a two-level weighting scheme. The methodological approach combines full-text reading and descriptive coding with a reproducible weighting scheme that accounts for mentions per study and integrates them at the corpus level using open-source tools. The results show a strong focus on Deep Learning (DL), with a greater presence in Digital Twins (DT) and BIM Modeling (BIMM); Multidimensional BIM (4D/5D) appears as a secondary line, while the Common Data Environment (CDE) and Clash detection (CD) are sporadic. The coupling of DL-DT and DL-BIMM predominates. Simultaneously, Machine Learning (ML) provides explainable analysis on structured data, and Generative Adversarial Networks (GAN) and Automated Machine Learning (AutoML) with Machine Learning Operations (MLOps) act as enablers for data generation/adaptation and deployment with traceability. It is concluded that advancing metrics and shared datasets, especially for CDE and CD, along with developing reproducible workflows oriented toward MLOps, are key to scaling AI in real-world environments.

1. Introduction

The Architecture, Engineering, Construction, and Operations (AECO) industry is under pressure to deliver safer, faster, and lower-carbon projects while integrating fragmented data across design, construction, and operations. In this context, BIM provides the information backbone, and AI supplies perception, prediction, and automation, together enabling near-real-time decisions at scale. However, current adoption remains uneven, and evidence is scattered. This review addresses that need by examining where AI is actually coupling with BIM in recent studies and the intensity of this coupling.

1.1. Context: Digital Transformation in AECO

The AECO industry is undergoing a sustained process of digitization and automation, aligned with the principles of Industry 4.0, which has been conceptualized as Construction 4.0 [1]. This transition drives the intensive use of data and the integration of technologies such as Building Information Modeling (BIM), the Internet of Things (IoT), advanced analytics, and Artificial Intelligence (AI), with the aim of improving coordination, production control, and asset operation [2]. Various recent reviews and frameworks indicate that this evolution depends on reliable information flows and the ability to close control loops almost in real time through measurements, automatic interpretations, and continuous feedback [3].

In this context, BIM acts as the informational backbone, while AI provides analytical, predictive, and prescriptive functions that enable new levels of automation and evidence-based decision-making [1]. The integration of these technologies enables the closure of the data, model, and decision cycle, enabling predictive analytics, such as energy optimization, and the orchestration of services on assets in near real-time, which is vital for efficiency in project management [2,3].

1.2. Fundamentals of BIM and AI

BIM is understood as a set of technologies, processes, and policies to generate, manage, and share digital representations of assets throughout their lifecycle [4]. Methodologically, BIM has been widely systematized in academic literature and international standards; for example, the ISO standards family 19650 defines principles and processes for information management, including the operation of a Common Data Environment (CDE) as an ‘agreed source of information’ for project teams [5].

An updated and sufficient framework to support the integration of BIM with AI is the ISO 19650 series, which defines principles and processes for information management and establishes the CDE as the ‘agreed source’ of data; achieving maturity in these processes and operating robust CDEs is a condition for reproducible, traceable, and auditable workflows in projects and assets [5,6].

Meanwhile, AI in AECO has shown rapid and heterogeneous advances: from Machine Learning (ML) and Deep Learning (DL) for classification, detection, and prediction, to generative models, such as Generative Adversarial Networks (GAN), for data synthesis and training set augmentation [3]. At the deployment frontier, Automated Machine Learning (AutoML) and Machine Learning Operations (MLOps) practices emerge as mechanisms to standardize, automate, and scale the model lifecycle in industrial environments [7]. Likewise, digital twins (DTs) represent a critical pathway to close the loop between operational data and analytical models, connecting representation, simulation, and control [2,8,9].

1.3. AI-BIM Knowledge Gap

Prior syntheses have mapped prominent AI applications in BIM and documented the rise in deep-learning pipelines for perception and forecasting in AECO, as well as the growing role of Digital Twins and information-management standards [3,5,8,10]. Focused surveys also describe generative methods for data augmentation and adaptation and outline emerging operational practices such as AutoML and MLOps [7,11,12,13]. However, a gap remains: existing reviews seldom align AI families with specific BIM application ambits under a standard, comparable metric; reporting is heterogeneous (tasks, datasets, outcomes), hindering cross-study comparison; CDE and clash-detection evidence is under-specified in terms of governance metrics; and lifecycle operationalization (AutoML/MLOps) is rarely quantified or linked to reproducible deployment.

This review aims to help close this gap by proposing a minimal common taxonomy of nine strategic ambits at the intersection of AI-BIM, five BIM applications, BIM Modeling (BIMM) 4D/5D, CDE, Clash Detection (CD), Multidimensional BIM 4D/5D (MBIM 4D/5D), (DT), and four families of AI (ML, DL, GAN, AutoML & MLOps), and by systematizing recent evidence (2022–2025) through a reproducible review protocol aligned with PRISMA 2020 [14].

Building on the gap identified above, the research gap will be addressed through a PRISMA-guided, two-level quantitative framework that standardizes heterogeneous findings into comparable scores: (i) model-class weights within each study and (ii) link weights between AI families and BIM application axes. This explicit bridge from gap to method yields decision-relevant rankings and reveals underexplored AI–BIM intersections that merit investment. Section 2 details the search strategy, eligibility criteria, and weighting procedure; Section 3 presents the comparative synthesis and key rankings.

1.4. Objective and Research Questions

The overall objective of this research is to characterize, through a systematic review of recent literature (2022–2025) following PRISMA 2020, the mechanisms of integrating artificial intelligence into the workflows and artifacts of BIM in the AECO sector. The analysis examines the principal AI families across key BIM applications to identify adoption and performance trends, document methodological and reporting gaps, and delineate opportunities for technology transfer and operational scaling.

Guided by the overall objective, the study structures its research questions for quantitative and qualitative analysis of the 2022–2025 corpus into two groups that drive the analysis. The first group, focused on AI methods and integration patterns, poses the following questions: which BIM flows and artifacts, BIMM, CDE, CD, MBIM 4D/5D, and DT, concentrate the highest integration with AI in the 2022–2025 literature, and with what relative intensity (e.g., standardized metrics like URW per study)? which AI families (ML, DL, GAN, and AutoML & MLOps) predominate in this integration, and for what purposes (perception/vision, prediction/time series, data synthesis, operational automation), and how do they vary according to the project lifecycle phase and context? The second group, focused on governance and reproducibility, proposes: what information management and governance requirements (e.g., ISO 19650, CDE practices, traceability/versioning, interoperability) influence the adoption, reproducibility, and scaling of AI-BIM in practice, including its operationalization through MLOps?

2. Materials and Methods

2.1. Context

A first tentative classification of the nine keywords shown in Table 1 was made to start bibliographical analysis based on authors such as Fujii et al. [15], Liu et al. [16], and Pan & Zhang [17], who have served as a basis to define the scope of the pillars that build the framework which characterize the relationship between AI and BIM. These keywords were then confirmed, as seen in Table 2. These descriptors have been selected for their ability to reflect, in an integrated manner, both the informational, parametric, and collaborative dimensions inherent to BIM, as well as the computational and algorithmic developments associated with various AI families. The combination of these categories enables a comprehensive analysis of the indexed literature from 2022 to 2025, focusing on domains that are driving substantial transformations in the AECO industry. The inclusion of each term is justified not only by its recurrence in high-impact publications but also by its potential to articulate convergent research lines and provide a coherent comparative framework for future academic explorations and professional applications.

Table 1. Strategic keywords for the integrated analysis of BIM applications and Artificial Intelligence (AI) technologies.

BIM Applications	AI Technologies
1. BIM Modeling (BIMM)	6. Machine Learning (ML)
2. Common Data Environment (CDE)	7. Deep Learning (DL)
3. Clash detection (CD)	8. Generative Adversarial Network (GAN)
4. Multidimensional BIM 4D/5D (MBIM 4D/5D)	9. Automated Machine Learning (AutoML) & Machine Learning Operations (MLOps)
5. Digital Twins (DT)

2.1.1. BIM Applications

This subsection defines the five BIM application ambits used for analysis: BIMM, CDE, CD, MBIM 4D/5D, and DT. These categories define where AI integrates with BIM and provide a basis for collecting evidence, comparing studies, and organizing results across the asset lifecycle.

• BIM Modeling: The BIMM is conceived as a digital, three-dimensional, and parametric representation of the asset, which integrates geometry, semantic metadata, and behavior rules throughout its lifecycle. This model serves as the core to coordinate design, information management, and activity coordination [4]. Its consideration as a keyword in this study is unavoidable, as it serves as the informational backbone upon which the main use cases and integration flows are based. The growing convergence of BIM and AI enables data-driven systems that address complex AECO problems more effectively [18].
• Common Data Environment: The CDE is conceived as the ‘source of information agreement’ for a project, that is, the collaborative environment where documents, models, and data are gathered, managed, and shared throughout the lifecycle, with traceability, version control, and standardized exchanges as guiding principles [5]. The UK BIM Framework specifies CDE workflows (states, approvals, metadata, automation) that improve consistency and transparency, while noting barriers such as maturity, tool integration, and governance [6]. In summary, the CDE serves as an information management infrastructure that enables reproducible and auditable flows and is a prerequisite for scaling AI-BIM integrations in real-world environments.
• Clash Detection: CD is defined as the automated process that identifies geometric or logical conflicts between model components to prevent problems before construction and reduce rework and costs. Recent literature incorporates the use of machine learning to filter out irrelevant clashes and prioritize those with greater impact, thereby decreasing false positives and coordination effort [19]. These investigations demonstrate that the combination of rule-based reasoning and supervised learning algorithms significantly enhances the classification and management of collisions in construction projects, leading to more efficient coordination.
• Multidimensional BIM 4D/5D: MBIM 4D/5D greatly enhances the planning and management of construction projects by incorporating time and cost factors into the digital model. BIM 4D incorporates time into the 3D representation, enabling construction sequence simulations and progress control, which allows for more effective visualization and better planning [20]. On the other hand, BIM 5D, which incorporates the dimension of costs, has proven to improve the accuracy of cost estimation and project management, overcoming barriers such as tool integration and staff training, which is crucial in complex environments [21]. MBIM links models to schedules and costs, improving planning and estimation in complex projects.
• Digital Twins: DTs are conceived as rich, semantic virtual representations of assets and processes that are bidirectionally synchronized with their physical counterparts through sensors, historization, and control systems, extending the scope of BIM into operation and maintenance. This BIM-DT integration enables closing the data-model-decision cycle, allowing for predictive analyses, such as degradation and failure, energy optimization, and service orchestration on the asset in near real-time. At the same time, it requires semantic alignment, interoperability, and an information management infrastructure to ensure traceability and governance [8,22,23].

2.1.2. AI Technologies

As mentioned in Table 1, this subsection introduces the four AI families considered in the review: ML, DL, GAN, and AutoML & MLOps. This set covers the main patterns of learning and data generation, as well as the operationalization of the model lifecycle. It will allow us to map, using comparable criteria, how each family aligns with the BIM ambits and the tasks for which they are used in the analyzed studies.

6.

Machine Learning: ML involves algorithmic methods that can learn patterns from large volumes of data to predict or classify without specifying rules in advance. In the AECO field, ML has established itself as an entry point for AI due to its versatility with structured and semi-structured data, with recurring applications in cost estimation, risk prediction, schedule optimization, as well as productivity and performance monitoring; literature mappings identify it as one of the most widespread families across the sector, along with computer vision and deep learning [3]. In particular, recent reviews on productivity document the systematic use of ML throughout the entire workflow, from data collection to modeling and evaluation, to improve estimation, monitoring, and decision-making on-site, reinforcing its role as an enabling technology in operational tasks, e.g., neural networks, Support Vector Machines, and DL approaches in scenarios related to construction labor and equipment productivity [24].

7.

Deep Learning: DL, as a subfield of machine learning based on deep neural networks, learns hierarchical representations of complex data (images, point clouds, video, and time series), enabling significant advances in computer vision, recognition, and forecasting relevant to AECO. In the construction industry, literature reviews show that DL has become a methodological pillar for tasks such as automatic progress tracking, site safety, defect detection/segmentation, activity recognition, and point cloud analysis, outperforming traditional approaches in various scenarios when sufficient labeled data is available [10]. These syntheses also emphasize that the performance of DL depends on the quality of data, consistent annotations, and reproducible pipeline design, as well as its integration with BIM workflows to close the data-model-decision cycle in real-world contexts [25].

8.

Generative Adversarial Networks: GANs consist of a generator and a discriminator that compete to produce realistic synthetic data and have become key tools for data augmentation/synthesis, domain adaptation, and improving robustness in typical AI-BIM tasks, e.g., semantic segmentation of images/point clouds, defect detection, closing gaps between simulated and real data. In the AECO field, a recent review in Automation in Construction systematizes their use cases and benefits in construction, including improvements in generalization when there is a scarcity of labels and heterogeneity among projects, and highlighting their cross-cutting potential to enable more reliable pipelines in AI-BIM integration [11]. Additionally, in civil engineering, a focused review reveals that GANs generate plausible data for training and reduce the domain shift between structural conditions, reinforcing their role as an enabling technology where real data is costly or difficult to obtain [12].

9.

Automated Machine Learning & Machine Learning Operations: AutoML automates key tasks in the modeling cycle, including algorithm selection, hyperparameter search, and pipeline design (including neural architecture search), thereby reducing dependence on manual expertise and standardizing experimentation. MLOps, on the other hand, applies DevOps principles to the model lifecycle for deployment, versioning, monitoring, and governance (model/data registry, traceability, drift control, Continuous Integration/Continuous Delivery). In the context of AI-BIM, the synergy between AutoML and MLOps is critical for transitioning prototypes into operational environments, such as CDE and DT, ensuring reproducibility, auditability, and maintainability of analytical artifacts throughout the asset’s lifecycle. Specialized literature synthesizes both the fundamentals and systems of AutoML as well as the architectures and best practices of MLOps for industrialization, providing a solid framework to scale AI-BIM workflows in a governed and sustainable manner [13].

Together, the nine keywords selected outline a thematic map that links the comprehensive digitalization of the asset lifecycle through enriched BIM models, CDE, and multidimensional analyses, with AI techniques that enhance automation, knowledge generation, and evidence-based decision-making. This shared taxonomy facilitates the identification of research gaps, guides the formulation of future questions, and provides the scientific community with a cohesive terminological reference for the interdisciplinary advancement of the AI-BIM agenda.

Below is Table 2. Summary of AI Technologies and BIM applications, which summarizes the main issues addressed in recent studies and the correspondence between AI techniques and BIM workflows. This synthesis illustrates how strategies such as ML, DL, GAN, and AutoML & MLOps practices are integrated in a complementary manner to enhance the efficiency and quality of processes throughout the digital asset lifecycle.

Table 2. Summarizes how AI families map to BIM applications/problems across recent studies.

AI/BIM	ML	DL	GAN	AutoML & MLOps	Citations
BIMM	Supervised learning on structured BIM attributes (e.g., quantity/attribute-driven estimation, classification).	Deep models for geometric/semantic feature extraction and recognition over images/point clouds tied to BIM.	Synthetic data to augment training for segmentation/classification of BIM-related imagery/point clouds.	Pipeline/search automation and governed experimentation for BIM-centric analytics (registries, versioning).	[4,10,16]
CDE	Prioritization/triage of metadata and documents aligned with ISO 19650 states; quality signals for governance.	Assisted validation for consistency checks (when applicable) under standardized exchange/workflow states.	Data synthesis to balance datasets for document/image-based tasks within governed repositories.	Model/data registries, CI/CD, lineage, and monitoring integrated with CDE processes (ISO 19650, UK guidance).	[5,6,7]
CD	Hybrid rule-based + ML to filter irrelevant clashes and prioritize high-impact conflicts.	3D/point-cloud perception and learned representations to support clash identification/triage when scans are available.	Scenario synthesis/augmentation to improve the robustness of detection pipelines in varied projects.	Continuous retraining/versioned datasets to adapt detectors across projects with traceable metrics.	[7,11,19]
MBIM 4D/5D	Prediction of schedule/cost deviations and parametric what-if analyses linked to the BIM/WBS.	Sequence-aware models for schedule/progress signals (when labeled timelines are available).	Synthetic schedules/sequences to stress-test plans and support robustness under uncertainty.	Comparative model selection and deployment as services for planning/control dashboards.	[4,20,21]
DT	Predictive analytics and supervisory control using operational data streams mapped to the asset model.	Perception/time-series modeling for state estimation, forecasting, and anomaly detection within the twin.	Stress-testing and domain adaptation (simulation-to-real) to improve twin robustness.	Lifecycle management (data/model versioning, monitoring) for deployed analytics in DT workflows.	[2,3,8]

Table 2. Summarizes how AI families map to BIM applications/problems across recent studies.

AI/BIM	ML	DL	GAN	AutoML & MLOps	Citations
BIMM	Supervised learning on structured BIM attributes (e.g., quantity/attribute-driven estimation, classification).	Deep models for geometric/semantic feature extraction and recognition over images/point clouds tied to BIM.	Synthetic data to augment training for segmentation/classification of BIM-related imagery/point clouds.	Pipeline/search automation and governed experimentation for BIM-centric analytics (registries, versioning).	[4,10,16]
CDE	Prioritization/triage of metadata and documents aligned with ISO 19650 states; quality signals for governance.	Assisted validation for consistency checks (when applicable) under standardized exchange/workflow states.	Data synthesis to balance datasets for document/image-based tasks within governed repositories.	Model/data registries, CI/CD, lineage, and monitoring integrated with CDE processes (ISO 19650, UK guidance).	[5,6,7]
CD	Hybrid rule-based + ML to filter irrelevant clashes and prioritize high-impact conflicts.	3D/point-cloud perception and learned representations to support clash identification/triage when scans are available.	Scenario synthesis/augmentation to improve the robustness of detection pipelines in varied projects.	Continuous retraining/versioned datasets to adapt detectors across projects with traceable metrics.	[7,11,19]
MBIM 4D/5D	Prediction of schedule/cost deviations and parametric what-if analyses linked to the BIM/WBS.	Sequence-aware models for schedule/progress signals (when labeled timelines are available).	Synthetic schedules/sequences to stress-test plans and support robustness under uncertainty.	Comparative model selection and deployment as services for planning/control dashboards.	[4,20,21]
DT	Predictive analytics and supervisory control using operational data streams mapped to the asset model.	Perception/time-series modeling for state estimation, forecasting, and anomaly detection within the twin.	Stress-testing and domain adaptation (simulation-to-real) to improve twin robustness.	Lifecycle management (data/model versioning, monitoring) for deployed analytics in DT workflows.	[2,3,8]

The matrix provides an initial, structured overview of how AI impacts key BIM concerns, serving as a baseline for the systematic analysis developed in this study. Building on this foundation, the review proceeds to examine empirical evidence, pinpoint methodological gaps, and outline opportunities for technology transfer to advance the AI-BIM agenda in the AECO industry.

2.2. Review Process

This review was conducted following the PRISMA 2020 guidelines [14]. The search was performed in the Web of Science (WoS) Core Collection—Index Expanded, enabling Enriched Cited References to maximize the retrieval of records with extended bibliographic traceability. The last search was conducted on 11 August 2025. The exact search string used was:

TS = ((“BIM Modeling” OR “Common Data Environment” OR “Clash detection” OR “Multidimensional BIM” OR “Digital Twin”)

AND (“Machine Learning” OR “Deep Learning” OR “Generative Adversarial Network” OR “AutoML” OR “MLOps”))

The inquiry was limited to the thematic areas of Engineering and Construction Building Technology, in line with the focus on AI + BIM in the building industry.

2.3. PRISMA Analysis

To ensure transparency and reproducibility, the PRISMA 2020 workflow for this AI-BIM review is reported in Figure 1. The process involves record search and consolidation, relevance and quality filtering, full-text evaluation using explicit criteria, and final inclusion decisions, structured across four stages: Identification, Screening, Eligibility, and Included. Each stage is detailed in the following subsections.

Figure 1. Flowchart of the PRISMA methodology for literature selection.

Screening at all stages was conducted independently by two experts, both of whom were experienced in BIM and AI; any disagreements were resolved by consensus, and no automation tools were utilized in this part of the review process.

2.3.1. Preliminary Identification

The keyword search was designed to capture, in parallel, the principal BIM integration vectors (BIMM, CDE, CD, MBIM, DT) and the most representative AI families (ML, DL, GAN, AutoML & MLOps). Beginning with 1212 articles, and after an initial retrieval in the WoS database and automatic deduplication, 59 unique records were identified. This set underpins the PRISMA flow and reflects the breadth of the AI-BIM spectrum over the study period.

2.3.2. Screening Process

Minimally restrictive methodological filters were used to preserve sensitivity while maintaining relevance:

• Years of publication (2022–2025). Seven documents were excluded (n = 7) because they fell outside the established time window, leaving 52 records. This temporal cutoff prioritizes recent publications in a rapidly evolving technological domain, ensuring that the findings reflect the current state of the art.
• Language: English. No exclusions were recorded (n = 0), maintaining 52 articles. The linguistic homogeneity facilitates synthesis and comparative evaluation, while also representing the publication standard in the field.
• Document type: Article. No exclusions were made (n = 0), so 52 was retained. The restriction to original or review articles ensures methodological rigor, availability of peer review, and completeness of metadata and results.

2.3.3. Eligibility Process

The 52 articles were evaluated in their entirety according to three explicit criteria: (i) journal prestige in the Journal Impact Factor (Q1), (ii) number of citations in the Web of Science as an indicator of influence and dissemination, and (iii) thematic relevance to a genuine intersection between AI and BIM. This combination of sources provided a balanced framework for judging quality, influence, and content. Based on this, 40 reports were excluded. These included those that did not meet the journal category criterion (JCR) (n = 17), had insufficient citations in the WoS (n = 22), and lacked a robust or explicit methodological integration between AI and BIM (n = 1).

2.3.4. Included Articles

After screening and eligibility assessment, 12 studies (n = 12) were included for in-depth analysis. Table 3 details each article, including its corresponding identifier, authors, DOI, and publication date.

Table 3. Included Articles (n = 12), 2022–2025, in the AI-BIM Review.

Number	Article Name	Authors	Publication Date
P001	Personal thermal comfort models using digital twins: Preference prediction with BIM-extracted spatial-temporal proximity data from Build2Vec.	Abdelrahman, MM; Chong, A; Miller, C	JAN 2022
P002	Construction Theory for a Building Intelligent Operation and Maintenance System Based on Digital Twins and Machine Learning	Zhao, YH; Wang, NQ; Liu, ZS; Mu, EY	FEB 2022
P003	A Digital Twin predictive maintenance framework of air handling units based on automatic fault detection and diagnostics	Hosamo, HH; Svennevig, PR; Svidt, K; Han, DG; Nielsen, HK	APR 2022
P004	Real-time mixed reality-based visual warning for construction workforce safety	Wu, SZ; Hou, L; Zhang, GM; Chen, HS	JUL 2022
P005	A Blockchain-Enabled Secure Digital Twin Framework for Early Botnet Detection in IIoT Environment	Salim, MM; Comivi, AK; Nurbek, T; Park, H; Park, JH	AUG 2022
P006	SpaRSE-BIM: Classification of IFC-based geometry via sparse convolutional neural networks	Emunds, C; Pauen, N; Richter, V; Frisch, J; van Treeck, C	AUG 2022
P007	Calibrating building simulation models using multi-source datasets and meta-learned Bayesian optimization	Zhan, SC; Wichern, G; Laughman, C; Chong, AD; Chakrabarty, A	SEP 2022
P008	Digital twin with Machine learning for predictive monitoring of CO2 equivalent from existing buildings	Arsiwala, A; Elghaish, F; Zoher, M	APR 2023
P009	Generating BIM Modeling from structural and architectural plans using Artificial Intelligence	Urbieta, M; Urbieta, M; Laborde, T; Villarreal, G; Rossi, G	NOV 2023
P010	Building information modeling-enabled multi-objective optimization for energy consumption parametric analysis in green buildings design using hybrid machine learning algorithms	Liu, Y; Li, TJ; Xu, WS; Wang, Q; Huang, H; He, BJ	DEC 2023
P011	A study of deep learning-based multi-horizon building energy forecasting	Ni, ZJ; Zhang, C; Karlsson, M; Gong, SF	JAN 2024
P012	Digital Management Methodology for Building Production Optimization through Digital Twin and Artificial Intelligence Integration	Piras, G; Muzi, F; Tiburcio, VA	JUL 2024

2.3.5. Risk of Bias and Language Considerations

This review conducts a non-meta-analytic mapping of evidence using a two-level comparative weighting (MCW/LW) approach, rather than estimating effect sizes; therefore, a formal assessment of the risk of bias at the study level does not apply to its objectives. This limitation is explicitly acknowledged, and transparency is emphasized through the protocolized search, independent double-checking, and reproducible coding and normalization.

Although the search was limited to English, the PRISMA flow indicates that this filter did not exclude any records (n = 52 before and after; articles excluded by language, n = 0), as shown in Figure 1. Furthermore, it is recognized that future updates should consider multilingual sources to broaden coverage.

2.4. Quantitative and Qualitative Analyses

A mixed-methods approach was employed to characterize the state-of-the-art of AI-BIM, as outlined in the corpus of 12 articles presented in Table 3. Quantitatively, weights for AI families and BIM subcriteria were estimated and subsequently integrated at the corpus level to obtain comparable indicators across studies. Qualitatively, problems, applications, tools, and methodological features were coded and triangulated with numerical patterns. The design follows the principles of multicriteria decision analysis (MCDA), articulating preferences and trade-offs among competing objectives [26].

2.4.1. Data Extraction and Initial Weight Processing

All analyses were conducted in Python (v.3.13.5) within Jupyter notebooks (v.7.3.2), relying exclusively on open-source software to ensure reproducibility. The 12 articles were ingested as PDFs, and text was extracted using pdfminer.six. Linguistic preprocessing employed spaCy (en_core_sci_sm) and PyTextRank, which comprised sentence and token segmentation, basic normalization (including case unification and removal of non-informative characters), and key-phrase extraction.

A seed dictionary for AI-BIM was constructed from the extracted text to guide the identification of substantive mentions. The AI dimension encompasses ML, DL, GAN, and AutoML & MLOps, while the BIM dimension includes BIMM, CD, MBIM 4D/5D, and DT. The dictionary incorporates common synonyms and acronyms and is applied case-insensitively to avoid formatting bias. Lemmatization is used solely to reinforce co-occurrence analysis and semantic consistency; it does not affect the frequency counts that underpin the weightings.

For each study, two input vectors are derived from the text counts: (i) a vector of AI families and (ii) a vector of BIM integration sub-criteria. When a study contains no explicit mention of any AI families, a neutrality rule assigns equal weights to all families to preserve comparability across studies. If no mentions are found for a given integration subcriterion, it receives a weight of zero in that study. These vectors support the subsequent steps: study-level normalization and corpus-level aggregation.

To aid readability, we define the symbols used in this subsection before formal expressions. We index AI families by i ∈ {ML, DL, GAN, AutoML/MLOps} and BIM sub-criteria by j ∈ {BIMM, CDE, CD, MBIM, DT}. Let n and m denote their counts, and N the number of papers. MCW refers to the study-level weight that summarizes the relative prominence of each AI family within a paper. LW denotes the within-family distribution that indicates which BIM application axes that family is directed toward in the study. GW captures the overall contribution of a specific AI family × BIM axis pair, reflecting the combined emphasis of the family and its focus on that axis in the study’s evidence. RW expresses this contribution in percentage terms to facilitate straightforward, like-for-like comparisons across rows and tables.

2.4.2. Calculation of Global Weights (GW) and Relative Weights (RW)

Two weights are defined at the study level: the model-class weight (MCW), which distributes weight across AI families within a study, and the link-weight (LW), which distributes weight across BIM integration subcriteria within the same study. These levels are combined via multiplicative aggregation to yield the global weight (GW) of each AI-BIM pair, as formalized in Equation (1):

$${GW}_{ij}={MCW}_i\times {LW}_j$$

(1)

where i goes through AI families and j goes through BIM sub-criteria.

A study-level normalization is established. To ensure comparability across articles with differing lengths and terminological densities, weights at each level are imposed to sum to 1 within each study, as shown in Equations (2) and (3):

$${\sum }_i^n{MCW}_i=1$$

(2)

where n is the number of AI criteria IA.

$${\sum }_j^m{LW}_j=1$$

(3)

where m is the number of BIM sub-criteria.

From Equations (1)–(3), it follows that the total global weights of the study also sum to 1, as seen in Equation (4):

$${\sum }_{i,j}^{n,m}{GW}_{i,j}=1$$

(4)

To facilitate interpretation, we express the overall weight as a percentage called relative weight (RW), as shown in Equation (5):

$${RW}_{i,j}=100\times {GW}_{i,j}$$

(5)

In the absence of mentions of AI families in a study (total count: zero), we apply a neutral rule assigning equal weights in MCW to avoid introducing arbitrary preferences. If a BIM subcriterion has no mentions, its component in LW is set to zero, so the corresponding pair does not contribute to GW or RW. These decisions preserve the normalization constraints.

Equations (2)–(4) prevent biases when specific terms do not appear in the text. Table 4 presents the pseudocode for the two-level weighting per paper, including inputs, fallback rules (uniform MCW and LW = 0), normalization of MCW/LW, and the computation of GW and RW, along with the per-study consistency checks.

Table 4. The proposed AI-BIM two-level weighting.

The Proposed AI-BIM Two-Level Weighting (Per Paper)
Input: text T, K_AI, K_BIM, recalc_weights
1. counts_ai ← freq_counts(T, K_AI)	# AI-family keyword counts
2. counts_sc ← freq_counts(T, K_BIM)	# BIM-subcriterion keyword counts
3. if recalc_weights and sum(counts_ai) > 0 then
4.  MCW ← normalize(counts_ai)	# sum(MCW) = 1 across AI families
5. else
6.  MCW ← uniform(|K_AI|)	# default equal weights
7. end if
8. if sum(counts_sc) > 0 then
9.  LW ← normalize(counts_sc)	# sum(LW) = 1 across BIM subcriteria
10. else
11.  LW ← zeros_like(counts_sc)	# no evidence for BIM subcriteria
12. end if
13. for each ai in K_AI do
14.  for each sc in K_BIM do
$15.\mathrm{GW}[\mathrm{ai},\mathrm{sc}]\leftarrow \mathrm{MCW}\left[\mathrm{ai}\right]\times$ LW[sc]
$16.\mathrm{RW}[\mathrm{ai},\mathrm{sc}]\leftarrow 100\times$ GW[ai, sc]
17.  end for
18. end for
19. Output: MCW, LW, GW, RW

2.4.3. Aggregation of Results and Unified Relative Weight (URW)

With the global weights by study, the results are integrated at the corpus level to obtain the unified relative weight (URW) for each BIM subcriterion. The GW_i,j of all studies are summed and projected onto each subcriterion j (summing over i) then normalized to 100%, as shown in Equation (6):

$${URW}_j=100\times {\displaystyle \frac{{\sum }_{p=1}^N{\sum }_i^n{GW}_{i,j}^{\left(p\right)}}{{\sum }_k^m\sum _{p=1}^N\sum _i^m{GW}_{i,k}^{\left(p\right)}}}$$

(6)

where N is the number of papers analyzed, in this case, 12 units.

Because weights sum uniformly per study, Equation (6) reduces to averaging GW by subcriterion across studies (including zeros), as outlined in Table 5.

Table 5. Corpus-level unified relative weight (URW).

Corpus-Level Unified Relative Weight (URW)
Input: {GW_p}_{p=1…N}	# per-paper GW matrices (AI×BIM)
1. M ← zeros(|K_AI|, |K_BIM|)
2. for p in 1…N do
3.  M ← M + GW_p
4. end for
5. sub_sum[j] ← Σ_i M[i, j]	# sum across AI families for each BIM subcriterion
$6.\mathrm{URW}\left[\mathrm{j}\right]\leftarrow 100\times$ sub_sum[j]/Σ_j sub_sum[j]
7. Output: URW	# percentages that sum to 100

2.4.4. Qualitative Analysis

A preliminary study was conducted to define the vocabulary of interest; drawing on a recent exploratory review, a set of keywords was identified to operationalize four AI families and five BIM application areas. This group guided both the search and selection of the 12-article corpus, as well as the mention-extraction procedure described in Section 2.1.

Each study was read in full and descriptively coded; we then triangulated these notes with MCW, LW, GW, RW, and URW to relate narratives to weights.

Finally, we compared corpus patterns with preliminary expectations to identify convergences, gaps, and shifts, thereby guiding the discussion without altering the calculations.

2.4.5. Quality Control and Numerical Consistency

The identities in Equations (2) and (3) are verified in each study; at the corpus level, ${\sum }_j^m{URW}_j=100\%$. In the absence of evidence in AI or BIM, the procedures described in Section 2.4.1 and Section 2.4.2 are applied, with equal weights across AI families and/or zero weights for BIM subcriteria without mentions, to maintain consistency and avoid biases due to terminological silence.

2.4.6. Export and Traceability

After computing the weights, three outputs are generated for downstream analysis and visualization:

• Per study tables with MCW, LW, GW, and RW.
• Corpus-level summaries of URW by AI family and by BIM subcriterion.
• The IA-BIM crossing matrix (URW, %) used in Section 3.6.

Traceability is ensured by keeping the study identifier (P00k) in every table and in the exported filenames, which provides a one-to-one link to the bibliographic record in Table 3 and allows any figure to be reconstructed from its source.

3. Results

This section presents the findings of the comparative analysis described in the previous Materials and Methods section.

3.1. Criteria Matrix AI-BIM

Table 6 presents the AI–BIM matrix. Columns are arranged in pairs: for each study P00k, the first column shows its Methodological Composite Weight (P00k_MCW), and the adjacent column shows its Result Weight (P00k_RW). Rows are organized in blocks. Each block begins with the AI family (ML, DL, GAN, AutoML/MLOps) and is followed by the five BIM application rows (BIMM, CDE, CD, MBIM, and DT). This paired-column layout, with rows grouped into blocks, makes the structure explicit and enables quick, like-for-like comparisons.

Table 6. (A) AI-BIM weighting results by (normalized MCW and RW)—Panel A: P001–P006. (B) AI-BIM weighting results by (normalized MCW and RW)—Panel B: P007–P012.

(A)
Criteria	P001_MCW	P001_RW	P002_MCW	P002_RW	P003_MCW	P003_RW	P004_MCW	P004_RW	P005_MCW	P005_RW	P006_MCW	P006_RW
ML	0.18		0.50		0.50		0.03		0.01		0.04
BIMM		9.92		4.58		9.84		0.50				3.14
CDE						1.04				0.06		0.24
CD
MBIM						1.55		1.17				0.73
DT		8.26		45.42		37.56		1.67		1.37
DL	0.41		0.25		0.42		0.90		0.94		0.96
BIMM		22.31		2.29		8.20		13.50				73.33
CDE						0.86				4.10		5.64
CD
MBIM						1.30		31.49				16.92
DT		18.60		22.71		31.30		44.99		90.19
GAN	0.36		0.25		0.07		0.03		0.03
BIMM		19.83		2.29		1.44		0.50
CDE						0.15				0.12
CD
MBIM						0.23		1.17
DT		16.53		22.71		5.48		1.67		2.73
AutoML & MLOps	0.05				0.01		0.03		0.01
BIMM		2.48				0.21		0.50
CDE						0.02				0.06
CD
MBIM						0.03		1.17
DT		2.07				0.78		1.67		1.37
(B)
Criteria	P007_MCW	P007_RW	P008_MCW	P008_RW	P009_MCW	P009_RW	P010_MCW	P010_RW	P011_MCW	P011_RW	P012_MCW	P012_RW
ML			0.13		0.03		0.49		0.02		0.04
BIMM				2.10		2.67		48.72				1.26
CDE												0.23
CD												0.04
MBIM										0.21		0.84
DT				10.80		0.36				1.69		1.41
DL	0.91		0.44		0.86		0.39		0.89		0.85
BIMM				7.10		76.11		38.46				28.30
CDE												5.15
CD												0.86
MBIM										9.85		18.87
DT		91.18		36.45		10.25				78.76		31.73
GAN	0.09		0.21		0.08		0.10		0.04		0.10
BIMM				3.42		6.68		10.26				3.46
CDE												0.63
CD												0.10
MBIM										0.49		2.31
DT		8.82		17.55		0.90				3.94		3.88
AutoML & MLOps			0.23		0.03		0.03		0.05		0.01
BIMM				3.68		2.67		2.56				0.31
CDE												0.06
CD												0.01
MBIM										0.56		0.21
DT				18.90		0.36				4.50		0.35

It should be noted that the empty cells shown in Table 6 (A and B) are equivalent to 0, as the study does not report information on the evaluated aspect. The sum of all RW values in a column, across all families and ambits, is approximately 100%, and the MCW values of the four families sum to 1 in each paper.

The above Table 6 (A and B) are recommended to be read in three complementary ways:

• Intra-paper (vertical): For each article, there are two columns in Table 6B: P00k_MCW and P00k_RW. The first column shows the relative weight of each AI family in a study, while the second column indicates the BIM axis to which the study focuses on that family.

Example 1.

For paper P007, column P007_MCW indicates that with an MCW of 0.91, DL is the predominant AI. Within this block, examining the second column, P007_RW, reveals that with an RW of 91.18%, the emphasis is on DT. This finding indicates that the DT pipeline is strongly driven by DL. The result is illustrated graphically in Figure 2.

Figure 2. E1-P007: DL to DT, DL-Driven Digital Twins (MCW = 0.91; DT RW = 91.18%).

Example 2.

When looking at paper P006 in Table 6A, column P006_MCW assigns a value of 0.96 to DL, and within that block, in column P006_RW, we observe that 73.33% is associated with the BIMM sub-criterion, which is consistent with the perception/model-centered workflows linked to the BIM model. The result is illustrated graphically in Figure 3.

Figure 3. E2-P006: DL to BIMM, Perception/Model Focus (MCW = 0.96; BIMM RW = 73.30%).

Example 3.

As shown in Table 6B, for paper P010, ML contributes an MCW of 0.49, and BIMM carries the emphasis with an RW of 48.72%, typical of analyses on structured BIM attributes. The result is illustrated graphically in Figure 4.

Figure 4. E3-P010: ML to BIMM, Structured Attribute Analytics (MCW = 0.49; BIMM RW = 48.72%).

• Intra-axis (horizontal, by sub-row): Based on the information shown in Table 6 (A and B), to see which AI family contributes more to a given BIM axis, select one sub-row and scan across studies within each AI block. In the DT subrow, the analysis from P001 to P012 reveals that the DL family consistently exhibits the presence of DT, with ML and GAN as secondary families, and AutoML/MLOps as occasional contributors.

Example 4.

In paper P007 (Table 6B), the weight RW = 91.18% of DT is concentrated in DL, with a negligible contribution, RW = 8.82%, from GAN; while in paper P008 (Table 6B), DL remains the main component (36.45%), but GAN (17.55%) and AutoML/MLOps (18.90%) provide significant support for DT. The result is illustrated graphically in Figure 5.

Figure 5. E4-DT Axis (P007 vs. P008): DL Main Component with GAN/AutoML Support.

• Between papers (horizontal by family row): An analysis of the MCW row of a given AI family reveals how its relevance varies across studies.

Example 5.

When analyzing Table 6 (A and B), in the DL row, the MCW is moderate from P001 to P003 with values between 0.25 and 0.42 but increases sharply from P004 to P007 (0.90 to 0.96) and remains high from P010 to P012 (0.85 to 0.89), indicating DL-centric workflows. Conversely, in the ML row, the strongest signals appear at P002 (0.50), P003 (0.50), and P010 (0.49), while ML is marginal from P004 to P007 (≤0.03), showing a shift away from ML in those studies. The result is illustrated graphically in Figure 6.

Figure 6. E5-MCW Across Papers: DL vs. ML Family Prominence.

3.2. Comparative Analysis of the Weights by Criterion for Each Paper

Figure 7 displays, for each paper (P001-P012), the MCW distribution across the four AI families (ML, DL, GAN, AutoML, and MLOps), with segments normalized and labeled for comparison. DL predominates in most studies, consistent with vision and time-series workflows on BIM artifacts and Digital Twins [2,8].

Figure 7. Distribution of MCW by Main Criterion in each Paper.

A second group of works presents hybrid ML-DL profiles, featuring a relatively balanced distribution between the two families. Notable examples of these approaches are projects P002 and P003, where ML achieves a feature mixing coefficient MCW of 0.50 and coexists with DL, which has coefficients of 0.25 and 0.42, respectively. ML handles structured attributes, while DL processes images, point clouds, and time series; both layers integrate within the BIM/DT flow [27].

Among the DL-centric profiles, several studies show an almost exclusive concentration; for example, P004 (MCW_DL = 0.91) and P005-P007 (values visually close to the upper end in Figure 7), which is consistent with vision and time-series workflows in BIM artifacts and digital twins [2].

A distinctive feature of the set is the GAN, present in P001 (MCW_GAN = 0.36) and P002 (0.25), and perceptible although in lesser preponderance in other cases. These contributions often serve as a complement or synthesis of data to strengthen deep learning models against label scarcity or domain changes between works and projects [27].

In terms of scalability and operation, the AutoML & MLOps family appears to be in its early stages in most of the studies shown in Figure 7 but gains greater prominence in study P008 (MCW_AutoML&MLOps = 0.21), where the lifecycle automation component acquires significant relative weight, signaling a shift from prototyping to governed deployment and monitoring [28].

Overall, the observed profiles reveal three behaviors: (i) DL predominates in perception and operation scenarios (DT, reality capture), (ii) hybrid ML+DL when tabular attributes and rich signals coexist, and (iii) occasional emergence of GAN/AutoML as accelerators, the former to address data limitations and the latter to professionalize the model lifecycle. The low systematic signal of AutoML & MLOps, as well as the heterogeneity in the use of CDE/ISO 19650, highlight areas of opportunity to scale the AI-BIM integration with greater traceability and reproducibility.

3.3. Aggregation and URW by Criterion

In this section, the corpus-level aggregation by AI families is presented. Table 7 summarizes the unified relative weight of each family (ML, DL, GAN, and AutoML & MLOps), computed from the per-study GW matrices and normalized according to Equations (5) and (6) so that the total equals 100%. This table enables a comparison, under a standard metric, of the average contribution of each family across the 12 studies.

Table 7. Unified relative weight (URW, %) by AI criteria.

AI Criterion	URW (%)
ML	16.45
DL	68.40
GAN	11.45
AutoML & MLOps	3.70
Total	100.00

Figure 8, based on the information in Table 7, is a normalized URW graph. It illustrates how the total weight is distributed among the five AI families (DL, ML, GAN, AutoML, and MLOps), with their proportions totaling 1, and each value is labeled for easy comparison. As shown in Figure 8, there is a clear bias towards DL, which accounts for 68.40% of the URW. This finding is consistent with the recent shift in the AECO sector toward perception tasks (2D-3D detection/segmentation, feature extraction) and time series analysis on BIM artifacts and Digital Twins, where DL performs better than purely tabular approaches. The added priority in the Digital Twin and BIM Modeling subcriteria is expected: these are the areas that benefit most from deep representations and high-frequency operational data [3].

Figure 8. Global distribution by AI Criterion (URW %).

Overall, three profiles emerge: DL-dominant pipelines for perception and operations (digital twins, reality capture); hybrid ML-DL when structured attributes and rich signals coexist; and tactical GAN-AutoML, where GAN mitigates data scarcity and AutoML & MLOps professionalize the model lifecycle. The weak and inconsistent signal of AutoML & MLOps, together with heterogeneous use of CDE/ISO 19650, points to clear opportunities to scale AI-BIM with stronger traceability and reproducibility [5,6,7].

3.4. Comparative Analysis of Sub-Criteria Weights for Each Paper

Figure 9 is a bar chart with one column per item (P001–P012), dividing the normalized weight of each item (LW, 0–1) across the five BIM sub-criteria (BIMM, CDE, CD, BIM, MBIM 4D/5D, and DT), with segment values labeled for easy comparison. A bimodal pattern is observed in the corpus, in which the weights tend to concentrate in DT and BIMM. In the first group, DT visually dominates in several studies, such as P003, P005-P006, P009, and P011–P012, which is consistent with flows oriented toward operation and monitoring, as well as the integration of sensor data or reality capture. This preference aligns with the literature linking digital twins to perception, diagnosis, and prognosis tasks supported by AI, especially when BIM artifacts are coupled with data streams for asset management and performance [8].

Figure 9. Distribution of LW by Main Subcriterion in each Paper.

The second pole corresponds to BIMM, with peaks at P001, P009, and, notably, P010, as shown in Figure 9. This emphasis suggests cases where the parametric model is central to the flow (extraction/management of quantities, semantic classification, and consistency validations), as well as scenarios where the structured attributes of the model are decisive, and data traceability takes precedence over unstructured signals [4,5].

The complementary sub-criteria appear with less weight and a selective nature. MBIM 4D/5D exhibits non-trivial contributions in P006, P009, and P012, which are compatible with planning and cost analyses linked to the model. CD mainly emerges in P006 and P012, indicating that, although coordination uses are consolidated in practice, they occupy a more limited methodological space when the focus of the study shifts toward DT or other operational ambits [2]. Meanwhile, CDE only records sporadic signals, with greater prominence at P010, suggesting that information governance and versioning are present to support the flow, but not as the primary analytical focus [5,6].

The combinatorial study reinforces three archetypes within the set: (i) central DT cases, focused on operational performance and continuous updating of the asset’s status; (ii) central BIMM cases, where the model and its structured attributes explain most of the weight; and (iii) hybrid cases that combine BIMM with 4D/5D or specific coordination (clash) to solve planning/execution problems. This qualitative interpretation, derived from full-text inspection and contrasted with the weights, provides explanatory context about why certain sub-criteria have relative importance and how they manifest in the reported flows.

3.5. Aggregation and URW by Subcriterion

The corpus-level aggregation by BIM subcriteria is presented in Table 8, which reports the unified relative weight (URW) for BIMM, CDE, CD, MBIM 4D/5D, and DT. The values are obtained by summing the GW matrices across the 12 studies, projecting onto each subcriterion, and normalizing to 100% using Equations (5) and (6).

Table 8. Unified relative weight (URW, %) by BIM subcriterion.

Subcriterion BIM	URW (%)
BIMM	34.38
CDE	1.53
CD	0.08
MBIM 4D/5D	7.43
DT	56.58
Total	100.00

Figure 10 shows a normalized horizontal bar chart (URW%) that ranks five BIM sub-criteria (BIMM, CDE, CD, MBIM 4D/5D, and DT) from highest to lowest, with percentages labeled at the end of each bar for direct comparison. A higher corpus concentration is observed in DT (56.58%), followed by BIMM (34.38%). The 4D/5D multidimensional MBIM has an intermediate share (7.43%), while CDE (1.53%) and CD (0.08%) are marginal.

Figure 10. Global distribution by Subcriterion BIM (URW %).

The dominance of DT and BIMM also aligns with evidence that deep learning and supervised learning deliver greater returns in perception and modeling tasks, such as vision on images/point clouds, and state prediction over operational data, than in traditional uses, such as clash coordination [3]. In contrast, the low weights for CDE and clash detection most likely reflect under-reporting and the lack of standardized metrics rather than lesser practical relevance. This interpretation aligns with the empirical profiling of BIM users, which reveals gaps in adoption and process standardization that obscure information-management practices, with CDE serving as the source of truth and a backbone for traceability [29].

3.6. Cross Between Criteria and Subcriteria

Table 9 is presented below, presenting the URW% between AI applications and BIM applications, which consolidates the AI-BIM matrix at the corpus level, based on the rows by AI families (ML, DL, GAN, AutoML, and MLOps) and the columns by BIM subcriteria (BIMM, CDE, CD, MBIM, DT). Each cell indicates the unified relative weight of the AI-BIM pair, obtained from Table 6 (A and B) (summing GW per study and normalizing to 100% according to Equations (5) and (6)). This matrix provides the numerical basis for subsequent visualizations, allowing for the interpretation of the dominant and secondary couplings in the corpus under a standard metric.

Table 9. URW (%) by AI family and BIM subcriterion.

Crossing AI-BIM	BIMM	CDE	CD	MBIM	DT	Total AI
ML	6.89	0.13	0.00	0.37	9.04	16.44
DL	22.47	1.31	0.07	6.54	38.01	68.40
GAN	3.99	0.08	0.01	0.35	7.02	11.45
AutoML & MLOps	1.03	0.01	0.00	0.16	2.50	3.70
Total BIM	34.38	1.53	0.08	7.42	56.57

The Sankey diagram in Figure 11, as well as the heatmap in Figure 12, shows that the flow mass primarily originates from DL and converges in DT and BIMM. This interaction suggests pipelines where DL processes rich signals, such as images, point clouds, and time series, and feeds digital twins for sensing, simulation, and prediction during operation. At the same time, BIM Modeling provides the semantic and geometric framework to align data and decisions. The direction from DL to DT and from DL to BIMM is consistent with the field’s evolution toward data-driven operational flows and with the role of DT as a computational container connected to the physical asset [3,8].

Figure 11. Sankey diagram. AI-BIM with global percentages (URW).

Figure 12. Heatmap of Average AI-BIM Integration.

The Sankey diagram also highlights the secondary contributions of ML to BIMM and DT, as well as the tactical role of GAN in data synthesis, augmentation, and simulation, which strengthens DL pipelines [11]. In contrast, the AutoML and MLOps streams refer to lifecycle orchestration and governance, including automated search of model pipelines, model and data registries, versioning, monitoring, and continuous integration and continuous delivery processes that automatically build, test, and deploy models safely into production, enabling traceability and reproducibility in DT/CDE workflows [7,13].

These patterns reflect multimodal architectures: DL addresses perception and asset status, while ML exploits structured model attributes for classification and estimation; both interact with BIM as semantic support for tasks and with the DT as an execution environment and feedback [8,25].

The weak couplings toward CDE and CD, with values close to zero across all families, suggest a reporting gap rather than a lack of relevance: the management of information and traceability enabled by the CDE is rarely quantified with comparable metrics. This scenario occurs even though the CDE acts as an authorized repository and single point of reference for data and models within the information governance framework. In turn, the still-nascent signal of AutoML & MLOps indicates that the integration of the model lifecycle (training, deployment, and monitoring) into the CDE is still under development, a necessary condition for scaling AI-BIM with reproducibility [7,30].

In line with the second analysis group, the corpus shows limited and heterogeneous reporting of ISO 19650/CDE artifacts and MLOps deliverables. When Digital Twin implementations are described, performance metrics dominate, while traceability and governance are rarely documented, including state transitions, version control, approvals, experiment tracking, and implementation and monitoring logs. This pattern helps explain the scarcity of comparable evidence on CDE and CD despite their practical relevance. The CDE is positioned as the contractual backbone and MLOps as the operational layer required to scale AI-BIM with traceability and auditability, in accordance with the International Organization for Standardization 2018 and with engineering guidance on maintainable machine-learning systems [5].

4. Discussion

This section consolidates the interpretation of the findings, first by contrasting the predefined conceptual map (AI families and BIM application axes) with the quantitative evidence (MCW, LW, GW/RW, URW) and then by synthesizing cross-cutting patterns, practical implications, limitations, and needs for future research.

4.1. Key Concepts Analysis

The nine keywords (BIMM, CDE, CD, MBIM 4D/5D, DT, DL, ML, GAN, and AutoML & MLOps) are contrasted with the analyses presented in Section 3. For each element, the expected role in the AI-BIM integration, the quantitative evidence from the study, and the implications and gaps are reported.

4.1.1. AI Technologies Analysis

The AI dimension is framed into four categories spanning the value cycle: ML for explainable analytics on structured data; DL for perception and representation in images, point clouds, and time series; GAN for data synthesis and adaptation; and AutoML & MLOps for operationalizing models (search, deployment, and monitoring).

• Machine Learning:

ML provides explainable and efficient models for tasks based on BIM attributes (classification, estimation, and control) and integrates with vision outputs or sensors when the input is tabular. In the corpus, it reaches URW = 16.43%, with weights concentrated in ML-DT = 9.04% and ML-BIMM = 6.89%, while MBIM = 0.37% and CDE with CD ≈ 0 are the residual components. The pattern suggests a complementary role to DL, which is useful when interpretability and data governance are required. It is advisable to standardize features derived from the model and establish comparable protocols between projects [3,24,31].

The prominence of ML-BIMM and ML-DT reflects contexts where project teams favor transparent models built on tabular BIM attributes and monitoring data for cost, delay, and failure risk predictions that must be explainable for contractual and managerial use [24,32].

• Deep Learning

DL addresses dense perception and representation across images, point clouds, and time series, and supports end-to-end pipelines within digital twins [25]. In our corpus, DL predominates (URW = 68.40%), with the strongest crossings in DL-DT (38.01%) and DL-BIMM (22.47%); the latter is exemplified by DL-based generation of BIM models directly from 2D plans [33]. An intermediate signal appears in DL-MBIM (6.54%), while CDE (1.31%) and CD (0.07%) remain low. This profile highlights DL’s reliance on labeled data and robust traceability, underscoring the need for standardized baselines and metrics to ensure reproducibility. In parallel, securing and auditing information flows in DT/CDE pipelines is crucial for trustworthy deployment [34].

The strong DL-DT and DL-BIMM couplings correspond to use cases where convolutional and sequence-based models absorb images, point clouds, and dense sensor streams anchored to BIM/DT, enabling automated defect detection, occupancy inference, and performance optimization in data-rich assets [10,22].

• Generative Adversarial Networks

Generative approaches play a crucial role in enhancement and synthesis, as well as domain adaptation, to strengthen perception models. In the corpus, for GANs, an URW of 11.45% is recorded, with focuses on GAN-DT at 7.02% and GAN-BIMM at 3.99%, with the rest being marginal. The evidence is consistent with scenarios of label scarcity or variability between works; however, a reporting gap persists. Clear protocols are necessary to evaluate the quality of synthetic data (fidelity, diversity, and bias) and to assess their impact on subsequent tasks reproducibly. This fact includes recording the origin and transformations of data and models, as well as maintaining version control for each delivery [11,27].

The limited GAN-BIMM/DT links indicate a focused but realistic application: synthetic scenarios and data augmentation, where labeled records are scarce, supporting the more robust training of inspection and monitoring models without altering the built asset [11,12].

• Automated Machine Learning & Machine Learning Operations

Low weight is not due to a lack of relevance, but to limited implementation and documentation: few works describe pipelines with version control for data and models, as well as service metrics. Integrating these elements within the CDE/DT is a condition for scaling with reproducibility [5,7].

The emerging connections between AutoML/MLOps and CDE, BIMM, and DT signal early efforts to formalize experiment tracking, deployment, and monitoring in governed environments, a prerequisite for scalable and auditable AI-BIM services in practice [6,15].

4.1.2. BIM Applications Analysis

This subsection addresses five BIM ambits (BIMM, CDE, CD, MBIM 4D/5D, and DT), where integration with AI is expressed. Quantitative patterns and qualitative evidence are synthesized to compare the role of each axis and identify implications and gaps.

• BIM Modeling

BIMM acts as a semantic-geometric backbone to structure AI data and support attribute extraction, labeling, and validation. It accounts for URW = 34.38%, primarily driven by DL = 22.47% and ML = 6.89%. Its guiding role is confirmed; however, there is a need for a greater explicit connection with CDE for traceability and with MBIM to close the planning-operation cycle [4,35].

The strong intersections with ML and DL demonstrate that many real-world implementations utilize BIMM as the stable, queryable layer from which quantities, states, and risks are computed, before extending to more complex DT-based control loops [10,22].

• Common Data Environment

The CDE should be the agreed-upon and traceable repository for data and model governance; however, its signal is weak (URW = 1.53%) and dispersed. This fact suggests a deficit in comparable metrics and reports, not a lack of importance. There is a need to operationalize within the CDE a provenance and traceability scheme (recording the origin and transformations of data and models), apply systematic versioning of datasets/code/models, and establish mechanisms for auditing analytical products (recording experiments and reproducible evidence), in line with ISO 19650 and the CDE guidelines [5,30,34].

• Clash Detection

As a coordination ambit, CD could benefit from supervised learning and graphs; however, it appears to be residual (URW = 0.08%). The focus of the corpus shifts toward perception/operation; there remains an opportunity for standardized datasets and tasks that measure the impact and prioritization of interferences with AI [4,19].

• Multidimensional BIM 4D/5D

MBIM 4D/5D links the model with time and cost to support planning, monitoring, and control, consolidating BIM as the semantic backbone of schedule and budget [4]. In the corpus, MBIM achieves a URW of 7.42%, with DL-MBIM accounting for 6.54%, while ML, GAN, and AutoML each contribute less than 1%. Functionally, MBIM acts as a hinge between modeling and operations; however, the scale of sequential learning in projects remains limited by inconsistent time series and labeling. These constraints are mitigated when MBIM is embedded in CDE/DT-centric digital management frameworks that integrate IoT and analytics to standardize WBS/Gantt, define control points, and produce reproducible temporal traces [36]. Complementarily, recent advances in 4D scheduling for modular construction reinforce the operational value of MBIM [21].

The moderate ML/DL–MBIM weights are consistent with applications where predictive models convert simulated 4D/5D scenarios into decisions on sequencing, logistics, and contingencies, improving schedule reliability and cost control beyond purely visual planning [21,24].

• Digital Twins

DT functions as an operational container for sensing, simulation, and prediction, and is the natural deployment destination for AI. It leads with URW = 56.57%, driven by DL-DT = 38.01% and, to a lesser extent, ML-DT = 9.04%, GAN-DT = 7.02%, and AutoML-DT = 2.50%. The domain confirms the shift toward operational exploitation and the need to consolidate DT-CDE-MLOps for traceability and continuous improvement [2,8,37].

Why does deep learning tend to dominate traditional learning? The leading DL-DT crossings suggest that twins become effective decision platforms when high-frequency sensing is combined with deep models for anomaly detection, predictive maintenance, and energy-control actions executed directly on the asset, rather than relying on static visualization tools [32,38].

4.2. Cross-Study Synthesis of Primary Couplings

Our findings generally confirm the suitability of the keywords defined in the preliminary phase, as well as four families of AI tools and five BIM applications to characterize the AI-BIM interaction in the corpus. The initial selection effectively guided the formation of the 12 articles set and the extraction of mentions. The nine-keyword scheme strikes a balance between precision and comprehensiveness; however, terminological variation in the literature (e.g., Building Information Modeling/Modeling, Digital Twin(s), Common Data Environment/CDE, 4D/5D planning/costing, conflict/coordination) can influence both retrieval and subsequent weighting. To mitigate bias, case-insensitive matching and controlled synonym lists were used.

Across heatmaps, Sankey flows, and crossing matrices, the interaction exhibits a coherent architecture: DL to DT and DL to BIMM carry the majority of the mass, while ML to DT and ML to BIMM provide a structured and interpretable layer. This pattern aligns with perception and time-series pipelines that are anchored to the asset model and complemented by those of feature-level analytics on BIM attributes [8,38]. Corpus-level averages, for example, DL to DT approximately 0.38 and DL to BIMM approximately 0.225, support a layered design in which DL addresses state estimation, forecasting, and anomaly detection within DT. In contrast, ML operates on semantically grounded BIM features, with BIM as the semantic backbone and DT as the operational container [3]. This reading aligns with field trends, including a BIM-enabled multi-objective design that couples simulation with random-forest metamodeling and NSGA-II search, and DT-based HVAC maintenance that integrates rule-based diagnostics with ML [16,32].

Conversely, the low signals observed in CDE and CD reflect measurement and reporting gaps rather than marginality. MBIM 4D/5D often acts as a bridge between modeling and operations, reinforcing the need to operationalize comparable governance indicators for CDE and CD without altering the taxonomy. Finally, GAN and AutoML/MLOps appear smaller but strategic; GAN as a lever for data synthesis and augmentation, and AutoML/MLOps as the path to governed operations with traceability—warranting the prioritization of studies that explicitly connect CDE, BIMM, and DT with these enablers to scale reproducibility [15,30].

In practice, generative approaches, such as GANs, remain underutilized due to practical barriers rather than limited potential: data scarcity and IP/privacy constraints, training instability with ambiguous evaluation, weak CDE-level provenance, and limited MLOps reporting, as well as high compute demands for 3D and sequential models. Addressing these with standardized datasets and metrics, explicit CDE lineage for synthetic data, and MLOps-ready assessment would enable disciplined, broader adoption and make GANs a more reliable complement to DL-centric pipelines.

In summary, the confrontation between what is defined in Section 2.1, where the keywords are justified, and the empirical findings confirms the core of the taxonomy: DT and BIMM as dominant applications, DL with ML as support, as analytical engines, and GAN, along with AutoML/MLOps, as data and operational enablers. The areas to strengthen do not require changing the categories but rather clarifying and applying comparable metrics for CDE and CD, as well as explicitly linking CDE-BIMM-DT with these enablers, to increase traceability, reproducibility, and scalability in AI-BIM integration.

4.3. Practical Implications

Looking ahead, the MCW/LW framework can be implemented by project lifecycle phase, maintaining standardization at the study level and aggregation at the corpus level (URW). This phase-based perspective reveals where AI families best integrate with BIM sub-criteria at each stage and how these patterns evolve, enabling direct comparisons under a transparent weighting scheme. Based on this evidence map, the implications translate AI-BIM signals into practical steps for AECO stakeholders. These include the need to adopt governance practices through robust CDE, operationalize reproducible workflows through MLOps, and prioritize use cases aligned with phases (e.g., DL to DT and ML to BIMM). Practical actions for owners, contractors/designers, operators, and PMO/procurement departments are then presented.

Based on the study’s findings, it is recommended that owners and public agencies formalize the CDE framework as the contractual backbone, including model/data registries, version control, procedure tracking, and consider MLOps deliverables, with deployment and monitoring records to maintain auditable and maintainable development technology DT. General contractors and designers align use cases with data reality, DL with DT for vision/time series, and ML with BIMM for attribute analysis, label planning, and documenting workflows for reproducibility. Facility/asset operators can establish results-based KPIs, such as energy usage and fault resolution time, linked to DT dashboards and institute retraining programs with rollback policies. Collectively, these steps translate the evidence into implementable policy, de-risking AI-BIM adoption. They ensure traceable, comparable pipelines with measurable cost, schedule, and sustainability impact.

4.4. Limitations and Future Research

Despite its contributions, this review has limitations. First, the results depend on a unique WoS corpus and the period from 2022 to 2025. Expanding the dataset could alter the relative weighting, particularly in poorly documented areas such as CDE and CD, by improving comprehensiveness and coverage. However, the workflow is fully reproducible and will be rerun with expanded datasets, recording changes in inclusion and classification. Second, the synthesis is non-meta-analytic and does not include a formal study-level risk-of-bias appraisal, consistent with our mapping and weighting aims rather than effect estimation. Third, the two-level weighting is derived from full-text, keyword-based mention counts, which can be influenced by reporting heterogeneity across studies. Finally, the inclusion thresholds, JCR quartile, and minimum citation count may skew the corpus toward more visible venues. Nevertheless, the framework is reproducible and can be strengthened in future updates by adding new databases (e.g., Scopus), incorporating multi-lingual sources, running sensitivity analyses on weights, and adopting standardized benchmarks to enhance comparability and coverage.

In terms of future research, three priorities are identified: (i) standardizing metrics and datasets for CDE and CD to enable evaluation of governance, traceability, and coordination under comparable criteria; (ii) strengthening data generation/adaptation (e.g., generative approaches or federated learning) in response to the scarcity and heterogeneity inherent in AI-BIM; and (iii) operationalizing AI within the CDE through MLOps practices (data/model versioning, automated testing, drift monitoring) and real-scale longitudinal assessments that quantify impacts on productivity, costs, energy, and emissions. These priorities follow directly from the empirical patterns observed in the corpus.

5. Conclusions

This study characterized the interaction between AI tools and BIM applications using a reproducible procedure, based on a systematic literature review. Full-text reading and descriptive coding were integrated with a two-level quantitative scheme, which assigns weights by AI family and BIM subcriteria and aggregates at the corpus level. This approach enabled the comparison of heterogeneous frameworks under a common metric basis (MCW, LW, GW, RW, URW). Normalizations were performed per study, and traceable export facilitated subsequent verification.

Taken together, the evidence indicates that research activity clusters around DT and BIMM because these tasks align with data-rich, model-centric workflows in which DL excels at perception and feature extraction, while ML supports attribute-level analytics. By contrast, CDE and CD appear underrepresented, not due to a lack of relevance, but rather because of fragmented reporting, missing shared protocols, and outcome measures that obscure the benefits of governance and integration. Generative approaches and AutoML-MLOps remain emergent and underdocumented, particularly in terms of data lineage, retraining, and deployment. The study highlights these asymmetries and outlines a practical agenda, consisting of consolidating evaluation resources for CDE and CD, aligning DL use cases with DT operational needs, and institutionalizing MLOps to move from prototypes to maintainable services.

In terms of contribution, a minimal taxonomy, five BIM ambits, and four AI families are provided, together with a standardized methodology for weighting and aggregation (MCW/LW/GW/RW/URW), enabling transparent comparisons between studies and facilitating joint interpretation of hybrid ML-DL profiles and emerging signals GAN, MLOps. The calculation structure and its default policies ensure consistency by assigning equal weights across AI families when mentions are missing, and by assigning zeros to BIM subcriteria when evidence is absent.

The findings underscore that governance and reproducibility are prerequisites for scalable AI-BIM. It is therefore recommended to operationalize ISO 19650 within the CDE through versioned information containers, explicit approval states, and documented lineage, and to institutionalize MLOps as standard deliverables, including registries, deployment logs, drift monitoring, and defined retraining and rollback procedures. Adopting this minimal indicator set will improve the identification of relevant texts, strengthen reproducibility, and increase the generalizability of the proposed weighting framework.

A further step involves operationalizing the framework as a phased, integrated dashboard that links the weights of the AI and BIM axes to impact metrics, such as energy consumption intensity and CO₂e for sustainability, schedule reliability and cost variance for 4D/5D, and fault resolution time for operations. Reassessing the weights by phase (design, construction, delivery, O&M) and combining them with these key performance indicators (KPIs) would generate comparable cost–carbon-performance ratios to guide decision-making and procurement. As the datasets expand, the same tools enable sensitivity analysis, allowing teams to track how AI and BIM contributions vary across phases and prioritize interventions with the highest verified value.