THE PARADIGM OF COGNITIVE COLLABORATION
BIM digitalization in AECO is evolving from simple automation to Augmented Intelligence. XAIBIM's purpose transcends repetitive tasks, focusing on an environment of augmented cognitive collaboration, where AI acts as a synergistic partner for professional transparent and reliable decision-making.
Problem: Semantic Fragmentation and Lack of Interoperability
The AECO sector suffers from data fragmentation and passive interoperability. XAIBIM Studio solves this by unifying knowledge into federated Knowledge Graphs (KG), transforming BIM models into lifecycle knowledge assets for automated and active decision-making.
Problem: Scarcity of High-Fidelity Data for AI
The lack of reliable datasets prevents the development of robust AI in construction. XAIBIM Studio addresses this by generating synthetic data assets that, once curated, form the CESAR (CESA Repository), the foundation for intelligent Digital Twins.
Problem: AI Opacity and Ineffective Normative Verification
'Black box' AIs are an unacceptable risk in construction. XAIBIM Nexus mitigates this risk with neuro-symbolic reasoning that enriches low-LOD models with GNNs and executes reliable and explainable deductive conformity verification.
Fundamental Doctoral Research Questions
Our research is structured around three thematic axes, each designed to address key challenges identified in the state of the art of AI applied to the AECO sector.
Axis 1: Curation and Expert Knowledge Foundations (XAIBIM Studio and CESAR)
This axis investigates the scientific methodology for creating an auditable, high-fidelity knowledge base, which is the fundamental prerequisite for any reliable AI system in the construction sector. The focus is on the curation process within XAIBIM Studio and the validation of the resulting CESAs.
- RQ1: What quantitative metrics (semantic consistency, information coverage) allow for validating the quality and auditability of a CESA (Curated, Explainable, Structured Asset) generated through the Human-in-the-Loop workflow of XAIBIM Studio?
- RQ2: How does the cryptographic integrity of the CESAR repository impact the robustness and bias mitigation of the XAIBIM Nexus engine trained exclusively on these assets?
- RQ3: What is the effectiveness of a semantic enrichment methodology in XAIBIM Studio to uniquely link cost information (5D) to BIM entities, ensuring a traceable Level of Information (LOI) in each CESA?
Axis 2: Engine Architecture and Validation (XAIBIM Nexus)
This axis focuses on the design, implementation, and empirical validation of the XAIBIM Nexus hybrid AI architecture. The research seeks to demonstrate the superiority of a neuro-symbolic approach for complex reasoning tasks in the AECO domain.
- RQ4: What is the performance (precision, F1-score) of the XAIBIM Nexus progressive neuro-symbolic pipeline (NLP→CNN→GNN→SWRL) in detecting constructability errors, compared to monolithic 'black box' AI models?
- RQ5: Can the GNN component of XAIBIM Nexus infer functional requirements implicit in low-LOD BIM models with validatable precision, semantically enriching CESAs to allow for early normative analysis?
- RQ6: How does the hybrid architecture of XAIBIM Nexus, which integrates contextual inference from GNNs with deductive reasoning from SWRL, allow for 4D/5D linking that is both automated and logically verifiable?
Axis 3: Measuring Explainability and Professional Impact (XAI)
This axis addresses the ultimate goal of the project: explainability (XAI). Research focuses on defining, measuring, and evaluating the impact of XAIBIM Nexus's transparent inferences on the trust and efficiency of construction professionals.
- RQ7: Which composite metrics model, combining source data traceability in the CESAR and clear logical inference chain in the Nexus, is most effective for quantifying the explainability of an AI recommendation in the AECO context?
- RQ8: How does an interface presenting XAIBIM Nexus inferences by visualizing relevant CESA sub-graphs and activated SWRL rules improve user trust and efficiency, compared to interfaces only showing final results?
- RQ9: What is the quantitative impact of the complete XAIBIM ecosystem (curation in Studio and analysis in Nexus) on reducing 4D/5D coordination errors, measured against traditional BIM workflows in a controlled case study?
A HOLISTIC METHODOLOGY FOR TRUSTWORTHY AI
A. Fundamental Principles of the Methodology
Our methodology is built on open interoperable standards (IFC, ISO 19650). We implement a Human-Centered AI (Human-in-the-Loop) model, where Explainable AI (XAI) enhances the construction professional's capability.
[18] C. Koch, et al., '"Explainable AI for engineering design: A unified approach of systems engineering and component-based deep learning demonstrated by energy-efficient building design,"' Advanced Engineering Informatics, 2024. DOI: 10.1016/j.aei.2024.102843
B. The Genesis of Trustworthy Knowledge
The flow starts in the XAIBIM Studio, where an expert, assisted by an RAG-LLM pipeline, curates and validates information on an ifcJSON artifact. The result is a CESA (Curated, Explainable, Structured Asset), which is cryptographically sealed and stored in the CESAR repository to ensure immutable trust.
[20] S. Kim, et al., '"Smart contract swarm and multi-branch structure for secure and efficient BIM versioning in blockchain-aided common data environment,"' Computers in Industry, 2023. DOI: 10.1016/j.compind.2023.103922
C. Applied Neuro-Symbolic Reasoning
The XAIBIM Nexus applies knowledge from the CESAR through a neuro-symbolic pipeline (NLP→CNN→GNN→SWRL). Our methodology optimizes the layers of these models to execute high-performance deductive reasoning for explainable conformity verification.
[22] F. F. Tafra, et al., '"Classification of architectural and MEP BIM objects for building performance evaluation,"' Advanced Engineering Informatics, 2024. DOI: 10.1016/j.aei.2024.102503
D. From Open Source to Cognitive Interface
We implement our methodology on an open-source technological stack (Blender, IFC.js, FastAPI). The interaction between the expert and the AI takes place through a multimodal cognitive interface designed for seamless collaboration.
[24] M. Nabizadeh, et al., '"Text-to-structure interpretation of user requests in BIM interaction,"' Autom. Constr., 2025. DOI: 10.1016/j.autcon.2025.106119
E. Validation and Metrics (KPIs)
The methodology's validation is carried out through Key Performance Indicators (KPIs) that measure both workflow impact and, centrally, the quantification of explainability for each Nexus inference.
[26] H. Son, et al., '"Enhanced model tree for quantifying output variances due to random data sampling: Productivity prediction applications,"' Autom. Constr., 2024. DOI: 10.1016/j.autcon.2023.105218
Validation Results and Metrics
XAIBIM System KPIs (Key Performance Indicators)
A. Knowledge Quality KPIs: Inter-annotator Consistency (Fleiss' Kappa - κ)
The reliability of the CESAR knowledge base is validated by measuring consistency among experts during the curation process in XAIBIM Studio. We use the Fleiss' Kappa (κ) coefficient to quantify the degree of agreement, ensuring the CESA repository is robust and standardized [27], [28].
κ = (P̄ - P̄ₑ) / (1 - P̄ₑ)CESAR
This KPI is fundamental to answering the research question on the quantitative validation of CESA quality. It measures the reproducibility and reliability of the expert curation process (Human-in-the-Loop).
[28] M. N. Johnson et al., '"Closing the artificial intelligence skills gap in construction: competency insights from a systematic review,"' Res. Innov. Eng., 2025. DOI: 10.1016/j.rineng.2025.106406.
B. Inference Engine Performance KPIs: Semantic Precision (F1-Score)
The performance of the XAIBIM Nexus in classifying BIM components is measured by the F1-Score. This metric evaluates the balance between precision and recall, quantifying the neuro-symbolic pipeline's capacity for semantic enrichment [29] and extraction of correct information [30].
F1 = (2 * Precision * Recall) / (Precision + Recall)XAIBIM Nexus
Directly measures the effectiveness of the Nexus hybrid architecture in correctly identifying and classifying BIM entities, a fundamental capability for subsequent conformity verification.
[30] X. Y. Zhang et al., '"Automatic bridge inspection database construction through hybrid information extraction and large language models,"' Digit. Build. Eng., 2024. DOI: 10.1016/j.dibe.2024.100549.
B. Inference Engine Performance KPIs: Verification Precision (Accuracy)
The effectiveness of the XAIBIM Nexus symbolic reasoning layer is evaluated with Conformity Verification Precision (CVP). This KPI quantifies the SWRL rule engine's capacity to correctly identify both compliance and non-compliance with regulations [31], [32].
CVP = (TP + TN) / (TP + TN + FP + FN)XAIBIM Nexus (Symbolic Reasoning Layer)
Validates the reliability of the Nexus deductive reasoning component, crucial for the promise of auditable AI and for mitigating regulatory risk.
[32] G. H. Green et al., '"BIM ontology for information management (BIM-OIM),"' J. Build. Eng., 2025. DOI: 10.1016/j.jobe.2025.112762.
C. Explainability and Impact KPIs: Explainability Quantification (SHAP)
To ensure auditable AI, the explainability of each XAIBIM Nexus inference is quantified using Feature Contribution Values (SHAP). This metric is key to overcoming AI adoption barriers in risk management [33] and building trust [34].
ϕᵢ(f,x) = Σ [fₓ(S∪{i}) - fₓ(S)] / N!XAIBIM Nexus (XAI Layer)
Directly answers the research question on defining metrics for explainability. It allows for objective measurement of model transparency.
[34] E. F. Brown et al., '"Applications of Explainable Artificial Intelligence (XAI) and interpretable Artificial Intelligence (AI) in smart buildings...,"' J. Build. Eng., 2025. DOI: 10.1016/j.jobe.2025.112542.
C. Explainability and Impact KPIs: Workflow Efficiency (Speedup)
The tangible impact of the XAIBIM ecosystem is measured by Computational Speedup. This KPI quantifies the efficiency gain in workflows compared to traditional manual methods [35], [36].
Speedup = T_manual / T_XAIBIMEcosistema XAIBIM (Studio + Nexus)
Validates the practical value and return on investment of the methodology, demonstrating its applicability and benefit for construction professionals.
[36] K. Torres, et al., '"Perceptions of the Influence of BIM Digital Models in Cost Overrun Management...,"' KSCE J. Civ. Eng., 2025. DOI: 10.1016/j.kscej.2025.100413.
D. Exploratory KPI: Generative Design Quality (FID)
As a future research line, XAIBIM Nexus's capability to generate design solutions will be evaluated using the Fréchet Inception Distance (FID). This KPI measures the quality and diversity of generated designs [37], [38].
FID(x,g) = ||μₓ - μg||² + Tr(Σₓ + Σg - 2(ΣₓΣg)¹/²)XAIBIM Nexus (Future Generative Capability)
Explores the natural evolution of neuro-symbolic reasoning toward generative capabilities, one of the frontiers of AI in AECO design.
[38] Q. R. Taylor et al., '"Generative AIBIM: An automatic and intelligent structural design pipeline integrating BIM and generative AI,"' Inf. Fusion, 2024. DOI: 10.1016/j.inffus.2024.102654.
Discussion: AI Implications and Governance in AECO (Architecture, Engineering, Construction & Operations)
The XAIBIM Discussion consolidates research findings into three governance pillars, demonstrating the viability of an auditable and explainable knowledge ecosystem for Trustworthy AI in a high-risk sector like AECO.
“For BIM models to reach their full potential in asset management, it is imperative to bridge the gap between graphical information and computable data through semantic enrichment. This process transforms the model into a robust digital twin, ensuring information consistency from initial capture to long-term operational management.”
“The incorporation of AI models into the built environment must prioritize explainability (XAI) and risk assessment, especially in high-risk systems where error can have catastrophic consequences.”
Knowledge Governance: From Data to 'Project Memory'
The XAIBIM ecosystem establishes robust data governance. Through the expert curation process in XAIBIM Studio [41], CESAs are generated to form the CESAR, an immutable and auditable project memory that guarantees algorithmic reliability and the long-term integrity of the Digital Twin [42].
Risk Mitigation: The Hybrid Neuro-Symbolic Paradigm
The XAIBIM Nexus mitigates the professional risk inherent in 'black box' AI. Its neuro-symbolic reasoning architecture uses GNNs to understand the topological context [43] and a SWRL formal verification engine to apply normative logic in a deterministic and auditable manner [44].
Improved Decision-Making: From Inference to Quantifiable Confidence
XAIBIM improves decision-making by going beyond simple inference. The methodology allows for confidence quantification through explainability metrics (XAI) like SHAP [45], providing the professional with an auditable safety level that validates and enhances their expert judgment in 4D/5D analysis [46].
Publications and Scientific Rigor
SCIENTIFIC FOUNDATION
Foundation for Creating a Trustworthy Repository
The creation of a trustworthy CESAR repository requires a rigorous curation process in the XAIBIM Studio. The methodology is founded on the ability to process unstructured information assets through deep learning [47] and organize them within an integrated semantic framework to ensure their logical consistency [48].
[47]: X. D. Wang, T. J. Li, et al., 'Reconstructing as-built beam bridge geometry from construction drawings using deep learning-based symbol pose estimation,' Adv. Eng. Inform., (2024). DOI: 10.1016/j.aei.2024.102808[48]: Y. C. Li, P. A. Zhao, et al., 'BIM-integrated semantic framework for construction waste quantification and optimisation,' Automat. Constr., (2024). DOI: 10.1016/j.autcon.2024.105842METHODOLOGICAL FRAMEWORK
Dual Methodology: Curation and Reasoning
The XAIBIM methodology is based on a strict separation of functions: the XAIBIM Studio handles knowledge curation and management throughout the asset's lifecycle [49], while the XAIBIM Nexus focuses exclusively on neuro-symbolic reasoning from this validated knowledge [50].
[49]: Q. M. Lu, J. L. Chen, et al., 'From design to operation: Multi-agent AI for virtual in-situ modeling of digital twins in BIM,' Automat. Constr., (2025). DOI: 10.1016/j.autcon.2025.106477[50]: A. B. Zhang, W. H. Li, et al., 'Automating the retrospective generation of As-is BIM models using machine learning,' Automat. Constr., (2023). DOI: 10.1016/j.autcon.2023.104937TECHNICAL ARCHITECTURE
The Neuro-Symbolic Hybrid Architecture of XAIBIM Nexus
The technical architecture of the XAIBIM Nexus validates the neuro-symbolic hybrid paradigm. A connectionist layer with Graph Neural Networks (GNNs) for topological analysis [51] is integrated with a symbolic layer using ontologies and logical rules (SWRL) for formal verification [52].
[51]: J. M. Chen, Q. B. Wang, et al., 'Graph Neural Networks for building and civil infrastructure operation and maintenance enhancement,' Adv. Eng. Inform., (2024). DOI: 10.1016/j.aei.2024.102868[52]: L. N. Li, K. M. Yu, et al., 'Artificial intelligence to enhance BIM-BEPS integration via IFC: Challenges, solutions, and future directions,' Adv. Eng. Inform., (2025). DOI: 10.1016/j.aei.2025.103824SCIENTIFIC CONTRIBUTIONS
Contribution to Explainable AI in Construction
Our scientific contribution establishes CESAR as the essential knowledge base (grounding) for reliable generative AI in construction [53]. This positions XAIBIM Nexus as an Explainable AI (XAI) tool designed for uncertainty management and robust optimization in civil engineering projects [54].
[53]: Y. Gao, G. Xiong, et al., 'Exploring bridge maintenance knowledge graph by leveraging GrapshSAGE and text encoding,' Automation in Construction, (2024). DOI: 10.1016/j.autcon.2024.105634[54]: L. M. Torres, G. R. Santos, et al., 'Multi-objective robust optimization for enhanced safety in large-diameter tunnel construction with interactive and explainable AI,' Reliab. Eng. Syst. Saf., (2023). DOI: 10.1016/j.ress.2023.109172OPEN SCIENCE
Repositories and Reproducibility
We guarantee reproducibility through open source. Using semantic rules (SPARQL) for logical verification in the Nexus [55] and developing open-source pipelines for data processing in the Studio [56] are examples of our commitment to open science.
[55]: H. P. Chen, S. L. Wu, et al., 'A Semantic Approach for Automated Rule Compliance Checking in Construction Industry,' IEEE Access, (2021). DOI: 10.1109/ACCESS.2021.3108226[56]: R. S. Jones, M. E. Davis, et al., 'Open-source automatic pipeline for efficient conversion of large-scale point clouds to IFC format,' Automat. Constr., (2025). DOI: 10.1016/j.autcon.2025.106303
Doctoral Forum: Debates and Frontiers of Trustworthy AI
XAIBIM proposes the transmutation of passive information assets into a computable and multidimensional knowledge ecosystem (from 2D to 10D). This project is articulated through three fundamental axes to address knowledge fragmentation in the AECO sector: Curator, Engine and Experience.
XAIBIM Studio Reliability: LLMs, Multimodal Extraction and Bias Risk
The XAIBIM Studio uses LLMs for knowledge extraction. The debate focuses on the fragility and bias of these models in the face of engineering jargon [57] and how Human-in-the-Loop control can, paradoxically, amplify bias if not quantified with consistency metrics like Fleiss' Kappa [58].
[58] '"Estimating demolition waste from residential interior photos: A Large Language Model solution,"' Automat. Constr., 2025. DOI: 10.1016/j.autcon.2025.106478.
XAIBIM Nexus Limits: Scalability of Neuro-Symbolic Reasoning
The hybrid paradigm (GNN/SWRL) of XAIBIM Nexus is the core of inference. The debate questions whether this synergy can scale to multidimensional analysis (4D/5D) [59] without introducing algorithmic fragility that compromises the reliability required of Trustworthy AI in construction [60].
[60] '"Trustworthy AI and robotics: Implications for the AEC industry,"' Automat. Constr., 2022. DOI: 10.1016/j.autcon.2022.104298.
XAI and Open Knowledge Responsibility
The transparency of XAIBIM Nexus is validated through explainability metrics (XAI) [61]. The ethical and professional debate focuses on the attribution of responsibility when the engine's logic is transparent and how access to the open knowledge of CESAR [62] impacts the validation of the engineer's judgment.
[62] '"Graph-based method for extracting spatial information from semi-formal text...,"' Automat. Constr., 2025. DOI: 10.1016/j.autcon.2025.106418.
Strategic Agenda and Vision
XAIBIM and European Obligatory Routes
The XAIBIM agenda is structured into a phased Digital Twin (DT) lifecycle [63], transitioning from data governance to professional Cognitive Augmentation. Development is rigorously iterative and structured [64], anchored in three fundamental axes: Curator (Phase I), Engine (Phase II), and Experience (Phase III).
[64]: M. A. Abuhussain, et al., 'Integrating Building Information Modeling (BIM) for optimal lifecycle management of complex structures,' Structures, (2024). DOI: 10.1016/j.istruc.2023.105831
ISO 19650
MANDATORY BIM IMPLEMENTATION (Critical Path)
REGULATED AI
EU AI Act and Data Governance
AGENDA 2030
SDGs and Infrastructure Resilience
AGENDA 2050
Multidimensional Inference and Climate Neutrality
FUTURE
Digital Twin and Cognitive Augmentation
MANDATORY BIM IMPLEMENTATION (Critical Path)
The XAIBIM project is a response to the BIM mandate in Europe. ISO 19650 requires information traceability and consistency, making the use of the Curator (Phase I) indispensable for structuring data (e.g., for BIM dimensions like 8D) [65]. The Engine, by formalizing logic, is the mandatory technical route for Automated Compliance Checking (ACC) of digital regulations, integrating semantic validation with engineering technical analysis [66].
EU AI Act and Data Governance
XAIBIM's design addresses data governance under the EU AI Act. By focusing on multimodal extraction [67], our system generates a training corpus that guarantees information reliability and traceability, mitigating algorithmic risks by employing ontology-based frameworks for formal regulatory verification [68].
SDGs and Infrastructure Resilience
XAIBIM contributes to SDGs 9 and 11 by providing the auditable data corpus (CESA) that is fundamental for reliable risk and resilience modeling [69]. The Engine implements risk modeling [70], demonstrating that sustainable infrastructure innovation is based on verifiable and dynamic Digital Twin information.
Multidimensional Inference and Climate Neutrality
Through its neuro-symbolic pipeline (Engine Axis), XAIBIM provides tools for accurate Life Cycle Assessment (LCA/LCC) of assets [71]. This is fundamental to meeting European Green Deal objectives, quantifying carbon footprint (6D) and optimizing design for climate neutrality through dynamic Digital Twins [72].
Digital Twin and Cognitive Augmentation
The agenda culminates with XAI integration (Experience Axis), positioning the system for the next generation of Digital Twins (DT) [73]. The goal is an auditable knowledge DT [74], where Cognitive Augmentation allows the professional to validate AI judgment in asset management.
The ifcJSON.XAIBIM Artifact: A Unified Knowledge Container
Explore the core of our methodology: the ifcJSON.XAIBIM artifact. It is not a simple file, but a knowledge container structured as a queryable knowledge graph [75]. Its digital twin hypermedia architecture [76] facilitates navigation across domains and ensures the traceability required by ISO 19650 and the AI Act [77].
[76]: S. Herlé & J. Blankenbach, '"Hypermedia-driven RESTful API for digital twins of the built environment,"' Automation in Construction, (2024). DOI: 10.1016/j.autcon.2024.105551
[77]: M. Huymajer, et al., '"Building Information Modeling in the execution phase of conventional tunneling projects,"' Tunnelling and Underground Space Technology, (2024). DOI: 10.1016/j.tust.2023.105539
SANDBOX (CONCEPT)
Semantic Enrichment Example
Selected Interface Controls:
> Create reinforced concrete beam C30/37 with 8m span, section 30x50cm...
{ "id": "3.2.1.1-BEAM", "ifc_version": "IFC4.3", "project_phase": "Execution", "LOD": "300", "LOI": "300", "bim_uses": [ "03 - Interdisciplinary Coordination", "05 - Cost Estimation (5D)", "06 - Structural Analysis" ], "expected_output": { "ifc_class": "IfcBeam", "object_type": "Reinforced Concrete Beam", "tag": "V-01", "geometry": { "profile": "Rectangular", "length": 8000, "width": 300, "depth": 500 }, "psets": { "Pset_BeamCommon": { "IsLoadBearing": true, "Span": 8000 }, "Pset_FireResistance": { "FireRating": "R 90" }, "Pset_EnvironmentalImpact": { "ThermalMass": 600 } }, "materials": [ { "name": "Concrete C30/37", "category": "Concrete" }, { "name": "Steel B500S", "category": "Steel" } ], "topological_relations": [ { "type": "IfcRelConnectsStructuralElement", "relating_element": "self", "related_elements": ["IfcColumn:P-01", "IfcColumn:P-02"] } ], "MEI": { "usage_context": "Structural analysis and execution planning.", "required_loi": 300 }, "MIDE": { "entities": ["IfcBeam", "IfcRelConnectsStructuralElement"], "parameters": ["IsLoadBearing", "Span", "FireRating", "ThermalMass"] }, "norms": [ { "code": "EN 1992-1-1", "purpose": "Structural requirements verification for concrete.", "jurisdiction": "EU" }, { "code": "CTE DB-SI", "purpose": "Fire resistance verification (R 90).", "jurisdiction": "ES-MD" }, { "code": "CTE DB-HE", "purpose": "Thermal inertia contribution.", "jurisdiction": "ES-MD" } ], "bim_dimensions": { "3D": "Geometric model for coordination.", "4D": "Linked to task 'Beam Concreting Level 1'.", "5D": "Concrete volume and steel weight for cost estimation.", "6D": "Contribution to energy efficiency and safety." } }, "xai": { "ai_type": "Hybrid (Neuro-Symbolic)", "explanation": "Inferred beam. Topological relations extracted. Regulatory validations applied: Eurocode 2, CTE DB-SI (Fire) and DB-HE (Energy).", "prov": { "wasGeneratedBy": "XAIBIM_Engine_v3.3", "generatedAtTime": "2025-09-19T19:45:00Z", "blockchain_hash": "0x456def7890ghi123jkl456mno789pqr123stu456vwx789yz0123456abc" }, "confidence": 0.99 }, "training_info": { "nlp_source": "The 'text_input' (not shown) and descriptions in 'psets' and 'object_type' train the NLP engine.", "cnn_source": "A 'snapshot_3d_id' (not shown) associated with this JSON trains the vision engine for beam recognition.", "gnn_source": "The 'topological_relations' block trains the GNN engine to learn structural connection patterns.", "symbolic_source": "The 'norms' and 'psets' blocks serve as the fact base for SWRL/ISO rule validation." } }
Immutable Visual Snapshot
Each ifcJSON artifact is linked to a 3D visual snapshot, an immutable representation sealed on the blockchain. This visual record is the training source for our AI models: through image processing and object detection techniques, CNNs learn to identify visual components [86], while GNNs analyze this data as non-Euclidean information to infer complex topological relationships in the BIM context [87].
[89]: '"Graph Neural Networks for building and civil infrastructure operation and maintenance enhancement,"' Advanced Engineering Informatics, (2024). DOI: 10.1016/j.aei.2024.102868
XAI Integration and Delta Synchronization
Explainability (XAI) is integrated as a semantic data layer to ensure auditability in cross-platform collaboration [78]. Change synchronization is managed through a differential transaction (delta.patch) method, ensuring an efficient and consistent update of CESAs [79].
[79]: F. Xue & W. Lu, '"A semantic differential transaction approach...,"' Automation in Construction, (2020). DOI: 10.1016/j.autcon.2020.103270
The Human Curator Role in the MLOps Cycle
The human expert oversees the MLOps (Machine Learning Operations) cycle, validating inferences to generate CESAs. This repository of ifcJSON artifacts [80] is the basis for supervised retraining and continuous improvement of XAIBIM Nexus [81].
[81]: Y. Mo & B. Li, '"ArchiWeb: A web platform for AI-driven early-stage architectural design,"' Frontiers of Architectural Research, (2025). DOI: 10.1016/j.foar.2025.06.002
Cryptographic Seal and Immutability Guarantee
Each CESA validated in the Studio is sealed by an immutable cryptographic record on a blockchain [82]. This process ensures the integrity and non-repudiation of the data making up the CESAR throughout the asset's life cycle [83].
[83]: W. van Groesen & P. Pauwels, '"Tracking prefabricated assets and compliance using quick response (QR) codes...,"' Automation in Construction, (2022). DOI: 10.1016/j.autcon.2022.104420
Knowledge Graph and Data Provenance (W3C PROV)
We transform BIM models into an AECO Knowledge Graph using Semantic Web technologies. To ensure auditability, we record the provenance of each piece of knowledge following the W3C PROV standard [84], allowing for explicit verification of each decision in the lifecycle [85].
[85]: J. Zhu, N. Nisbet, et al., '"Releasing the power of graph for building information discovery,"' Automation in Construction, (2025). DOI: 10.1016/j.autcon.2025.106034
Standard Evolution and IFC5 Preparation
The XAIBIM architecture is designed for standard evolution, prepared to integrate construction process (as-built) data [86]. Our graph-based approach is capable of modeling the complex topological relationships that will emerge in future standards like IFC5 [87].
[87]: C. Emunds, J. Frisch, & C. van Treeck, '"Spatial Link Prediction: Learning topological relationships in MEP systems,"' Advanced Engineering Informatics, (2025). DOI: 10.1016/j.aei.2025.103414
Integration with Blender and Open Source Ecosystem
Our workflow integrates with leading open-source tools like Blender, not just as a visualization platform, but as an active component in the data lifecycle and AI validation.
Blender as a Synthetic Data and Validation Engine
We integrate Blender into the XAIBIM Studio MLOps flow as an engine for procedural synthetic data generation. We use its API for data augmentation and employ it as a validation environment to mitigate geometric "hallucinations" of LLMs in parametric CAD workflows.
[89]: '"The status, evolution, and future challenges of multimodal large language models (LLMs) in parametric CAD,"' Expert Systems with Applications, (2025). DOI: 10.1016/j.eswa.2025.127520
Interoperability and Semantic Validation
The integration is based on open interoperability (IfcOpenShell). To ensure CESA quality, we use the SHACL constraint language for ifcOWL ontology validation, allowing for semantic verification of complex technical requirements directly in the modeling environment.
[91]: '"Validation of technical requirements for a BIM model using semantic web technologies,"' Adv. Eng. Inform., (2024). DOI: 10.1016/j.aei.2024.102426
The Knowledge Graph and AI Agents
In Blender, we transform static BIM into a dynamic Semantic Connectivity Graph (IFC-Graph). This data structure allows XAIBIM Nexus to operate as a Multi-Agent System (MAS) and execute advanced logic, such as route calculation based on semantic relations.
[93]: '"Design of a Multi-Agent System for exploiting the communicating concrete in a SHM/BIM context,"' IFAC-PapersOnLine, (2020). DOI: 10.1016/j.ifacol.2020.11.061
The Studio-Nexus Cycle and Open Science
The Blender integration materializes the MLOps cycle: XAIBIM Studio generates the CESAR, an ifcJSON.XAIBIM repository that functions as a Graph-Based BIM (GBIM). XAIBIM Nexus is trained on this corpus. By publishing the CESAR, we contribute to open science for XAI research in construction.
[95]: '"Explainable artificial intelligence (XAI): Precepts, models, and opportunities for research in construction,"' Adv. Eng. Inform., (2023). DOI: 10.1016/j.aei.2023.102024
Stack Tecnológico del Frontend
Next.js y TypeScript
La plataforma se construye con Next.js como framework para Gemelos Digitales, diseñado para la integración de datos de múltiples fuentes [96]. Utilizamos TypeScript para la gobernanza del código y para garantizar la integridad estructural del artefacto ifcJSON.XAIBIM en todo el ecosistema [97].
- Persistencia Científica: Su capacidad de renderizado híbrido (SSR/SSG) asegura que los resultados publicados sean inmutables y accesibles, un requisito para la reproducibilidad de la investigación.
- Gobernanza de la Lógica: Las "Server Actions" encapsulan la lógica de negocio, creando un pipeline de datos auditable desde la interfaz hasta el backend.
- Integración de Datos: Permite la integración de datos de múltiples fuentes, fundamental para un Gemelo Digital que debe consolidar información de BIM, IoT y otras bases de datos.
[97]: S. Sepasgozar, et al., '"BIM-driven transformation of waste management toward enhanced reduction and circularity in the built environment,"' Waste Management, (2025). DOI: 10.1016/j.wasman.2025.115105
Three.js e IFC.js
El stack de visualización 3D con Three.js e IFC.js es la interfaz Human-in-the-Loop del XAIBIM Studio. Actúa como la capa de interacción para la retroalimentación visual en tiempo real [98], permitiendo al experto validar el enriquecimiento semántico derivado de procesos como la Comprobación de Relaciones Geométricas [99].
- Interfaz de Curación: Sirve como la herramienta human-in-the-loop para auditar visualmente los resultados del "Curator".
- Bucle de Retroalimentación: Constituye la capa de interacción de nuestro marco de integración físico-digital, un componente esencial para el feedback loop en tiempo real.
- Transparencia del Enriquecimiento: Hace visible el resultado del enriquecimiento semántico basado en la comprobación de relaciones geométricas.
[99]: G. Lilis, et al., '"BIM-based semantic enrichment and knowledge graph generation via geometric relation checking,"' Automation in Construction, (2025). DOI: 10.1016/j.autcon.2025.106081
Interacción Humano-Céntrica
La interfaz de usuario se construye con Tailwind CSS y Shadcn/ui para implementar un marco de trabajo centrado en el humano que garantiza la claridad en la interacción con el Gemelo Digital [100]. El sistema facilita una colaboración humano-robot eficaz entre los expertos y los agentes de IA del Nexus [101].
- Comunicación Científica Clara: Un sistema de UI consistente garantiza que la presentación de datos y métricas de XAI sea no ambigua.
- Interacción Humano-Céntrica: Implementa un marco de trabajo centrado en el humano que facilita la colaboración entre expertos, usuarios y agentes de IA.
- Accesibilidad y Colaboración: El cumplimiento de estándares como WCAG es fundamental en la colaboración humano-robot.
[101]: '"Biosignal measurement for human-robot collaboration in construction: A systematic review,"' Advanced Engineering Informatics, (2025). DOI: 10.1016/j.aei.2025.103652
Zustand para la Reproducibilidad
Utilizamos Zustand para la gestión del estado de la aplicación. Su arquitectura minimalista es fundamental para la estandarización y eficiencia en flujos de trabajo colaborativos de BIM [102] y, crucialmente, para garantizar la reproducibilidad científica de cada sesión de curación en el XAIBIM Studio [103].
- Eficiencia y Estandarización: Un gestor de estado como Zustand contribuye a la estandarización de datos y la eficiencia del sistema.
- Base para Middleware: La capacidad de Zustand para integrarse con middlewares alinea nuestra arquitectura con este enfoque para la interoperabilidad.
- Reproducibilidad Experimental: Permite guardar, compartir y restaurar el estado exacto de una sesión, un requisito para la reproducibilidad científica.
[103]: '"Semantic enrichment for BIM-based building energy performance simulations...,"' Journal of Building Engineering, (2024). DOI: 10.1016/j.jobe.2024.110312
Backend Technology Stack: The Dual AI Architecture
🧠 XAIBIM Studio Backend
Purpose: Knowledge creation. A pipeline to generate the Audited Semantic Training Corpus (CESA).
LLMs
XAIBIM Studio implements an LLM-based agent functioning as a Cognitive Assistant, allowing professionals to perform "3D spatial relationship queries" through natural language [100], an approach aligned with "technology acceptance" factors in the AECO sector [101].
- Cognitive Assistant: The AI agent assists the expert in curation, interpreting complex natural language queries.
- Intuitive Spatial Query: Resolves the challenge of extracting topological information from IFC models through a conversational interface.
- Evidence-Based Adoption: The design is based on empirical factors determining technology acceptance in the sector.
[101]: 'An interactive system for 3D spatial relationship query by integrating tree-based element indexing and LLM-based agent,' Advanced Engineering Informatics, (2025). DOI: 10.1016/j.aei.2025.103375
RAG
To mitigate "hallucinations," the Studio uses Retrieval-Augmented Generation (RAG). This enhanced knowledge framework [102] grounds LLM proposals in a verified corpus, improving the "quality and reproducibility" of generated CESAs [103].
- Evidence Grounding: Drastically reduces LLM "hallucinations" by forcing it to base its answers on reference documents.
- Quantitative Improvement: Increases the accuracy and reliability of extracted data, a requirement for high-risk systems.
- Native Traceability: Allows every piece of generated data to be intrinsically linked to its original documentary source.
[103]: 'Generative artificial intelligence in construction: A Delphi approach, framework, and case study,' Alexandria Engineering Journal, (2024). DOI: 10.1016/j.aej.2024.12.079
Blockchain
To guarantee CESAR integrity, we integrate Digital Twin and Blockchain for lifecycle management [111]. Blockchain is used to ensure the security, integrity, and transparency of each validated CESA [112].
- Immutable Record: Every expert validation is cryptographically sealed, creating a tamper-proof history.
- Decentralized Governance: Trust resides not in a central authority, but in cryptographic consensus.
- DT-BC Synergy: Enables a secure and reliable data flow between the physical asset and its validation record.
[112]: 'Integrating Digital Twin and Blockchain for dynamic building Life Cycle Sustainability Assessment,' Journal of Building Engineering, (2024). DOI: 10.1016/j.jobe.2024.111018
⚙️ XAIBIM Nexus Backend
Purpose: Knowledge application. A 'white box' Hybrid AI for semantic enrichment of BIM models.
Python/FastAPI
The ecosystem operates on a Python and FastAPI backend for its robustness in the scientific ecosystem. This architecture supports BIM workflow automation through its API [96] and the implementation of real-time predictive models [97].
- Scientific Ecosystem: Native access to cutting-edge libraries like PyTorch and Transformers.
- High Performance: FastAPI's asynchronous nature is ideal for managing AI operations without blocking the system.
- Predictive Intelligence: The architecture is designed to support not only command execution but also learning from user interaction.
[97]: 'Predictive modeling: BIM command recommendation based on large-scale usage logs,' Advanced Engineering Informatics, (2025). DOI: 10.1016/j.aei.2025.103574
MLOps and Containers
XAIBIM is built on a Service-Oriented Architecture (SoA) [98]. Each AI component is packaged in Docker containers and orchestrated through microservices [99], guaranteeing a reproducible and scalable MLOps lifecycle.
- Guaranteed Reproducibility: Docker containerization eliminates dependency issues and ensures a consistent environment.
- Scalability and Resilience: Microservices architecture allows each component to scale independently.
- CI/CD: Facilitates automation pipelines for testing and deploying new versions of AI models.
[99]: 'A Semantic Digital Twin for the Dynamic Scheduling of Industry 4.0-based Production of Precast Concrete Elements,' Advanced Engineering Informatics, (2024). DOI: 10.1016/j.aei.2024.102677
NLP, CNN, GNN, SWRL, XAI
XAIBIM Nexus executes a pipeline combining NLP with Transformers (SBERT) for semantic analysis [104], Computer Vision (Mask R-CNN) for instance segmentation [105], Topological Analysis with GNNs to understand relational context [106], and Symbolic Logic with Semantic Web (ifcOWL) for deterministic validation [107], culminating in an XAI layer (LIME) to ensure transparency [108].
- Multimodal Analysis: Processes and correlates information from texts, 2D images, and 3D models.
- Contextual Reasoning: GNNs allow AI to understand relationships between model components.
- Auditable Validation: Symbolic logic layer provides deterministic and transparent rule checking.
[105]: 'Generating BIM model from structural and architectural plans using Artificial Intelligence,' J. Build. Eng., (2023). DOI: 10.1016/j.jobe.2023.107672
[106]: 'Integrating generative and parametric design with BIM...,' Applied Sciences, (2025). DOI: 10.1016/j.apples.2025.100253
[107]: 'Knowledge-enhanced ontology-to-vector for automated ontology concept enrichment in BIM,' J. Ind. Inf. Integr., (2025). DOI: 10.1016/j.jii.2025.100836
[108]: 'A LIME-LSTSNM approach based green building sustainability prediction using BIM design,' Sustain. Comput. Inform. Syst., (2025). DOI: 10.1016/j.suscom.2025.101155
Specialized Databases
The data architecture relies on polyglot persistence. We employ Graph Databases (Neo4j) for BIM topological representation [109], Document Databases (MongoDB) for ifcJSON.XAIBIM artifacts, and Relational Databases (PostgreSQL) for transactional operational data [110].
- Optimized Storage: Each data type is stored in a system designed for maximum efficiency.
- High-Performance Queries: Allows complex topological queries in the graph database.
- Scalability and Flexibility: The architecture allows each database system to scale independently.
[110]: 'AI-driven integration of digital twins and blockchain for smart building management systems...,' Journal of Building Engineering, (2025). DOI: 10.1016/j.jobe.2025.112439
Support Infrastructure and Governance
Security, Governance, and Regulatory Compliance
The XAIBIM platform is built on a security and governance model that prioritizes human-centered innovation and regulatory compliance. We implement standard authentication flows like OAuth 2.0/JWT and use Web3 technologies to protect intellectual property. Our framework proactively addresses cybersecurity concerns and ethical AI principles, ensuring adherence to regulations such as GDPR and the EU AI Act.
Modern Authentication: Uses open standards to manage user sessions and securely protect API access.
Intellectual Property Protection: The architecture is designed to protect the confidentiality of commercial data and the intellectual property of BIM models.
[76] '"Trustworthy Artificial Intelligence in AECO: A review of ethical and legal implications,"' Eng. Const. Archit. Manag., 2024. DOI: 10.1108/ECAM-05-2023-0453.
AI Life Cycle (MLOps)
XAIBIM adopts an MLOps methodological framework for the full life cycle management of our AI models. We implement a Human-in-the-Loop (HITL) cycle where expert validation from the "Curator" feeds back into the continuous retraining of the "Engine". Our architecture separates NLP model training from its deployment on an inference server, ensuring a robust, auditable, and constantly improving process.
Continuous Improvement: The system is not static; it learns and improves continuously from expert interaction and validation.
Model Life Cycle Management: Automates the process of retraining, validation, deployment, and monitoring of AI models.
[78] '"Human-in-the-loop machine learning: A state of the art review,"' Knowl. Based Syst., 2024. DOI: 10.1016/j.knosys.2024.111823.
Human Validation (Human-in-the-Loop - HITL)
Our methodology focuses on a Human-in-the-Loop (HITL) cycle, where expert knowledge is the final authority validating the CESA. This process transforms validation into automated safety checking that not only corrects but actively improves the system's intelligence. This approach is a pillar of our data governance strategy.
Expert Validation as Source of Truth: The professional's tacit knowledge is captured and formalized as a quality standard.
Active Feedback Loop: Every expert action feeds the MLOps cycle for continuous "Engine" retraining.
[80] '"Continuous learning architectures for infrastructure asset management,"' Resour. Conserv. Recycl., 2025. DOI: 10.1016/j.resconrec.2025.107982.
Governance and Standards (ISO 19650 & AI Act)
The entire XAIBIM methodology operates under a strict governance framework, aligned with ISO 19650 for BIM information management and the principles of the EU AI Act. This approach ensures compliance with the highest interoperability standards and aligns with future digitalization and automation trends.
Compliant Information Management: The workflow is compatible with ISO 19650 principles.
Trustworthy AI by Design: The system meets EU AI Act requirements for high-risk systems.
[82] '"Regulatory compliance of AI in the European Union: The AI Act and its impact on infrastructure,"' Polit. Soc., 2023. DOI: 10.1177/00323292231189212.