AI-Driven IFC Processing for Automated IBS Scoring

Abstract

The assessment of Industrialized Building System (IBS) adoption in construction projects—a critical metric for evaluating prefabrication levels and construction modernization—remains largely manual, time-intensive, and prone to inconsistencies, with practitioners typically requiring 4–8 h to evaluate a single building using spreadsheet-based frameworks and visual documentation review. This paper presents a novel AI-enhanced workflow architecture that automates IBS scoring through systematic processing of Industry Foundation Classes (IFC) building information models—the first documented integration of web-based IFC processing, visual workflow automation (n8n), and large language model (LLM) reasoning specifically for construction industrialization assessment. The proposed system integrates a web-based frontend for IFC file upload and configuration, an n8n workflow automation backend orchestrating data transformation pipelines, and an Azure OpenAI-powered scoring engine (GPT-4o-mini and GPT-5-0-mini) that applies Construction Industry Standard (CIS) 18:2023 rules to extracted building data. Experimental validation across 136 diverse IFC building models (ranging from 0.01 MB to 136.26 MB) achieved a 100% processing success rate with a median processing duration of 61.62 s per model, representing approximately 99% time reduction compared to conventional manual assessment requiring 4–8 h of expert practitioner effort. The system demonstrated consistent scoring performance with IBS scores ranging from 31.24 to 100.00 points (mean 37.14, SD 8.84), while GPT-5-0-mini exhibited 71% faster inference (mean 23.4 s) compared to GPT-4o-mini (mean 80.2 s) with no significant scoring divergence, validating prompt engineering robustness across model generations. Processing efficiency scales approximately linearly with file size (0.67 s per megabyte), enabling real-time design feedback and portfolio-scale batch processing previously infeasible with manual methods. Unlike prior rule-based compliance checking systems requiring extensive manual programming, this approach leverages LLM semantic reasoning to interpret ambiguous construction classifications while maintaining deterministic scoring through structured prompt engineering. The system addresses key interoperability challenges in IFC data heterogeneity while maintaining traceability and compliance with established scoring methodologies. This research establishes a replicable architectural pattern for BIM-AI integration in construction analytics and positions LLM-enhanced IFC processing as a practical, accessible approach for industrialization evaluation that democratizes advanced assessment capabilities through open-source workflow automation technologies.

1. Introduction

The global construction industry faces mounting pressure to improve productivity, sustainability, and quality outcomes through industrialization strategies [1,2]. Industrialized Building Systems (IBS)—defined as construction systems in which components are manufactured in a controlled environment, transported to the site, and assembled into structures with minimal additional site work [3,4]—offer significant potential for addressing these challenges through prefabrication, modular construction, and systematic assembly processes.

However, accurate quantification of IBS adoption levels within building projects remains predominantly manual, requiring expert practitioners to analyze architectural documentation, extract component specifications, and perform calculations through spreadsheet-based frameworks [5]. This manual approach typically requires 4–8 h of expert effort per building, limiting the scalability of IBS assessment for large portfolios and iterative design optimization.

The widespread adoption of Building Information Modeling (BIM) technologies, particularly the Industry Foundation Classes (IFC) open data scheme, presents unprecedented opportunities for automating construction analytics [1,2,6]. IFC models encapsulate comprehensive geometric representations alongside rich semantic metadata describing building elements, material specifications, spatial relationships, and functional classifications. Yet, despite this data-rich environment, the translation from IFC model content to quantified IBS scoring metrics remains largely unexplored in both academic literature and professional practice [7,8].

1.1. Limitations of Current IBS Scoring Methodologies

Current IBS scoring methodologies face several critical limitations that impede their effectiveness and scalability: Time-Intensive Manual Extraction. The manual extraction of component counts, dimensional parameters, and classification categories from architectural documentation requires substantial time investment. A typical comprehensive IBS assessment for a medium-sized building project requires 4–8 h of expert analysis [4,5], representing a significant resource allocation that many organizations struggle to justify for routine project evaluation. Inconsistent Human Interpretation. Human interpretation of drawings and specifications introduces significant variability in classification decisions, calculation errors, and inconsistent application of scoring criteria across different evaluators. Research by Blismas et al. [3] and Kamar et al. [4] documented inter-rater reliability coefficients as low as 0.65–0.75 for manual IBS assessments, indicating substantial disagreement even among experienced practitioners when categorizing construction methods and computing industrialization metrics. This variability undermines the credibility of assessment outcomes and complicates comparative benchmarking across projects or organizations Scalability Constraints. Scenario analysis, iterative design optimization, and portfolio-wide assessments involving hundreds of building models cannot be efficiently accommodated through conventional manual methods [9,10]. This limitation prevents systematic industrialization strategy optimization during design development phases when design changes are most cost-effective and impactful. Data Duplication and Traceability Issues. Information already encoded within BIM models must be manually re-entered into separate assessment frameworks, creating unnecessary duplication and synchronization challenges that increase error risk and reduce process efficiency [11,12]. Additionally, spreadsheet-based calculations often lack comprehensive audit trails linking final scores to specific building elements. When assessment results are questioned or disputed, reconstructing the logical chain from building components through classification decisions to final scores becomes prohibitively difficult.

1.2. Research Objectives

This paper addresses these limitations by developing and demonstrating an integrated workflow architecture that combines web-based user interfaces, workflow automation platforms, and artificial intelligence technologies to automate IBS scoring from IFC building models. This research pursues six interconnected research objectives: System Architecture Development: Develop a complete system architecture integrating frontend interface, backend workflow orchestration, and AI-powered scoring logic for end-to-end IFC-to-IBS transformation. Technical Implementation: Demonstrate a working implementation using contemporary technologies, including HTML5/JavaScript frontend, n8n workflow automation, and Azure OpenAI language models, to validate technical feasibility. Data Extraction Methods: Implement methods for extracting, classifying, and quantifying building elements from IFC models, addressing inherent data heterogeneity and attribute incompleteness challenges. AI Integration: Leverage large language model capabilities for interpreting complex construction classification CIS 18:2023 (Construction Industry Standard for IBS Content Scoring) [5] criteria to structured building data. Performance Assessment: Assess the system’s processing efficiency, consistency, accuracy, and scalability compared to conventional manual assessment approaches. Theoretical Contribution: Derive theoretical insights regarding the integration of BIM, AI, and workflow automation for the construction industry’s digital transformation.

1.3. Research Contributions

This research makes several distinct contributions to construction informatics:

Novel Integration Pattern: First documented integration combining web-based IFC processing, n8n visual workflow automation, and LLM-powered scoring engines for automated IBS assessment, establishing a replicable architectural pattern for construction analytics applications. Methodological Contribution: A formal methodology for translating regulatory scoring frameworks (CIS 18:2023) into LLM-executable prompt specifications through structured prompt engineering, temperature control strategies, and validation protocols. Practical Implementation: A functioning system with real IFC model processing capabilities, validating technical feasibility and identifying implementation considerations that emerge through actual deployment [9,10]. LLM-Based Rule Interpretation: Effective encoding of domain-specific rules within AI agent prompts for consistent automated interpretation, advancing beyond rule-based systems requiring extensive manual programming [13,14]. Accessible Technology Stack: Open-source workflow automation (n8n) and standard web technologies promoting accessible digital transformation without vendor lock-in. Portfolio-Scale Processing: Batch processing capability enabling portfolio-wide industrialization assessment previously infeasible with manual methods. The subsequent sections systematically develop this workflow architecture through literature synthesis, establishing theoretical foundations (Section 2), detailed methodology and implementation description (Section 3), results presentation and discussion of theoretical and practical implications (Section 4), and conclusions regarding contributions and future research directions (Section 5).

2. Literature Review

2.1. Key Terminology and Definitions

This section introduces essential terminology used throughout this paper to ensure accessibility for readers from diverse backgrounds: Building Information Modeling (BIM): A digital representation process involving the generation and management of digital representations of physical and functional characteristics of buildings. BIM serves as a shared knowledge resource for information about a facility, forming a reliable basis for decisions during its lifecycle [1]. Industry Foundation Classes (IFC): An open, international standard (ISO 16739-1:2024) [5] for exchanging BIM data between different software applications. IFC provides a neutral data schema enabling interoperability across heterogeneous software platforms [6]. Industrialized Building System (IBS): Construction systems characterized by mass production of building components in factories, transportation to construction sites, and assembly into buildings with minimal wet work. IBS emphasizes prefabrication, modular construction, and systematic assembly processes [3,4]. The IBS concept and its associated scoring framework originate from the Malaysian construction industry, where the Construction Industry Development Board (CIDB) formalized IBS classification and content scoring to promote industrialization and reduce dependence on unskilled labor [5]. While the term “IBS” is specific to the Malaysian regulatory context, analogous concepts exist internationally under different nomenclature—including offsite construction (OSC) and modern methods of construction (MMC) in the United Kingdom [3], prefabricated prefinished volumetric construction (PPVC) in Singapore, and industrialized construction in Scandinavian countries—all sharing the core principle of shifting construction activities from on-site manual processes to controlled factory environments [4]. The CIS 18:2023 scoring framework used in this research represents the most mature and quantitatively formalized assessment methodology among these international approaches, providing explicit numerical scoring tables that map construction system characteristics to industrialization metrics specificity that makes it particularly amenable to computational automation. Certified IBS assessors are construction professionals who have completed formal training and examination administered by CIDB Malaysia, qualifying them to evaluate building projects against the IBS Content Scoring criteria; certification requires demonstrated competence in classifying structural systems, wall systems, and other construction solutions according to the prescribed factor tables, and typically demands a background in civil engineering, architecture, or quantity surveying with practical construction industry experience [5]. CIS 18:2023: The Construction Industry Standard 18:2023 published by the Construction Industry Development Board (CIDB) Malaysia, providing the normative framework for calculating IBS Content Scores based on structural systems, wall systems, and other simplified construction solutions [5]. n8n: An open-source workflow automation platform enabling visual programming of data transformation pipelines through node-based interfaces, increasingly adopted for business process automation [15].

2.2. Building Information Modeling and IFC Data Schema Evolution

Building Information Modeling has evolved from basic geometric representation systems into a comprehensive information management framework encompassing structured project data across multiple dimensions and lifecycle phases [1,2,6]. While BIM provides foundational data structures that can support digital twin implementations, it is important to distinguish BIM as an information management methodology from digital twins, which represent a distinct use case involving real-time synchronization between physical assets and their virtual counterparts. The Industry Foundation Classes specification, developed and maintained by buildingSMART International, provides an open, neutral data schema enabling interoperability across heterogeneous software platforms throughout building lifecycles. IFC’s EXPRESS-based schema defines hierarchical entity structures representing:

Physical building components (IfcBuildingElement subclasses such as IfcWall, IfcSlab, IfcBeam, and IfcColumn). Spatial organization (IfcSpace and IfcBuildingStorey). Material specifications (IfcMaterial and IfcMaterialLayerSet). Relationships connecting entities through semantic associations (IfcRelAggregates and IfcRelContainedInSpatialStructure).

This comprehensive semantic framework theoretically enables automated extraction of construction-relevant parameters, including component enumeration, dimensional quantification, material classification, assembly relationships, and custom attributes [16,17].

However, practical IFC model implementations exhibit significant heterogeneity arising from software vendor interpretations, modeling practice variability, schema version fragmentation (IFC2x[M15]3, IFC4, and IFC4.3), and incomplete semantic enrichment [7,8]. These challenges necessitate robust processing strategies that accommodate data quality variations while maintaining analytical reliability.

2.3. BIM-Based Automated Construction Analytics

It is essential to recognize that BIM constitutes an information management framework supported by a series of standards spanning from Information Requirements through to Delivery, Verification, and Acceptance of information containers—rather than being a data repository itself [18,19]. The utility of IFC models for automated analytics depends critically on whether the required information was specified in the Exchange Information Requirements (EIR) and subsequently modeled to the appropriate Level of Information Need (LOIN). The progression from manual document review to automated BIM data extraction represents a fundamental shift in construction information management practices, contingent upon appropriate information specification and delivery.

2.3.1. Quantity Takeoff and Cost Estimation

In the domain of quantity takeoff and cost estimation, multiple studies demonstrate automated extraction of material quantities from IFC models for cost estimation purposes [20,21]. It should be noted that IFC models only contain data suitable for quantity takeoff and cost estimation when specifically modeled to the required level of information need as defined in project information requirements. These systems parse IFC entities to extract dimensional properties such as volumes, areas, and linear measures, and then apply unit cost databases to generate preliminary cost estimates. However, while successful in reducing quantity surveying effort when appropriate data is available, these systems often struggle with IFC data quality variations and require substantial validation against traditional takeoff methods to establish confidence.

2.3.2. Automated Code Compliance Checking

Automated building code verification systems analyze IFC models against regulatory requirements using rule-based reasoning engines, enabling systematic compliance verification at scale. Early work by Eastman et al. [13] established foundational approaches for encoding building codes as computable rules. Subsequent research has expanded these capabilities to address complex multi-clause regulations and performance-based code provisions [17,22]. However, the semantic gap between natural language code text and formal computational logic remains a significant challenge requiring substantial expert effort to translate regulations into machine-readable formats.

Recent work has begun addressing this semantic gap through natural language processing. Zhang and El-Gohary [23] developed NLP-based information extraction pipelines for regulatory documents, achieving F1 scores of 0.84–0.91 for requirement extraction. Zheng et al. [14] advanced this further with deep learning approaches that integrate knowledge graphs with BIM rule checking, demonstrating 87% accuracy in automated compliance verification.

However, these systems still require substantial training data curation and domain-specific model fine-tuning. The emergence of foundation models like GPT-4 and subsequent versions offers an alternative paradigm where general-purpose reasoning capabilities can be directed toward construction domain tasks through careful prompt engineering, potentially reducing the development effort barrier that has limited the adoption of automated compliance systems. Zhang et al. [24] recently demonstrated that LLM-based code interpretation achieves comparable accuracy to rule-based systems (89% vs. 91%) while requiring only 15% of the implementation effort.

2.3.3. Energy and Performance Simulation

IFC-to-simulation model transformation workflows enable automated energy analysis by extracting building geometry, material thermal properties, and HVAC specifications [25]. Similar approaches have been applied to daylight analysis, computational fluid dynamics, and acoustic simulation. Recent AI-enhanced performance prediction has achieved R2 values of 0.94 for energy consumption prediction [26], with transformer-based architectures outperforming traditional simulation in computational efficiency by factors of 100–1000× [27].

2.3.4. Construction Scheduling and Structural Analysis

IFC models have been integrated with construction scheduling methodologies through automated activity sequence generation based on spatial and logical dependencies between building elements [28,29]. By analyzing geometric relationships and construction logic encoded in BIM models, researchers generate precedence networks for project scheduling. Linking 3D geometric models with 4D schedule information has demonstrated substantial improvements in constructability analysis and logistics planning. Workflows extracting structural element geometries, material properties, and loading conditions from IFC models for finite element analysis have been demonstrated in multiple contexts. These approaches enable structural engineers to leverage BIM data for detailed analysis without duplicate modeling effort.

2.4. Artificial Intelligence Applications in Construction

The application of artificial intelligence technologies in construction has accelerated dramatically with the emergence of large language models and generative AI systems, opening new possibilities for addressing longstanding challenges in construction information processing and decision support [30].

2.4.1. Document Processing and Information Extraction

LLMs demonstrate remarkable capabilities for extracting structured information from unstructured construction documents, specifications, and technical drawings [31]. GPT-4 achieves 91% accuracy in construction specification interpretation tasks, significantly outperforming previous BERT-based models (78% accuracy) without domain-specific fine-tuning [32]. Wu et al. [33] developed a GPT-4-based system for automated construction contract review that identified 94% of critical clauses compared to 67% by junior legal reviewers.

2.4.2. Machine Learning for Decision Support

Machine learning models provide recommendations for construction method selection, material choices, and scheduling optimization based on project characteristics [34]. Deep learning architectures have enabled automated defect detection, progress monitoring, and safety compliance verification from site imagery [35,36]. Recent multimodal AI advances integrate visual perception with natural language reasoning, achieving 88% accuracy in identifying safety hazards while providing natural language explanations [37].

2.4.3. Gap Analysis: BIM-AI Integration for Assessment Automation

Integration of AI capabilities with structured BIM data processing for automated assessment applications remains an emerging research area with limited documented implementations.

Table 1. Comparative analysis of BIM-based automated assessment approaches.

Approach	Rule Encoding	Semantic Handling	Implementation Effort	Adaptability
Traditional Rule-Based [17]	Manual programming	Limited	High (months)	Low
NLP-Enhanced [19]	Semi-automated	Moderate	Medium (weeks)	Medium
Deep Learning [22]	Training-based	Good	Medium-High	Medium
Present research (LLM + Workflow)	Prompt-based	High	Low (days)	High

summarizes the comparative analysis of existing approaches against the present research.

2.5. Research Gaps and Research Positioning

Four significant gaps in the literature motivate the current research: Limited IBS Assessment Automation: Systematic approaches to classifying and quantifying industrialization indicators from IFC models remain substantially underdeveloped [3,4]. It is important to note that, despite the existence of automated BIM-based tools for adjacent domains such as code compliance checking [13], quantity takeoff [20,21], and energy simulation [25], no automated or semi-automated system for IBS Content Scoring has been documented in the academic literature or commercial software landscape as of the time of writing. Existing IBS assessment workflows rely entirely on manual procedures: certified assessors visually inspect architectural documentation, manually extract component counts and dimensional parameters, consult CIS 18 factor tables, and compute weighted scores using spreadsheet templates provided by CIDB Malaysia [5]. This absence of automation is partly attributable to the specialized and regionally specific nature of IBS scoring—rooted in the Malaysian construction regulatory framework—which has limited the commercial incentive for software vendors to develop dedicated tools, and partly to the interpretive complexity of construction system classification decisions that resist straightforward rule-based encoding. Disconnected AI and BIM Research: Integrated systems combining structured IFC data processing with AI reasoning capabilities are not well documented, missing opportunities for synergistic integration [38,39]. Underexplored Workflow Automation: Modern workflow automation platforms for construction analytics remain largely unexplored despite widespread adoption in other industries [40]. Theory-Practice Divide: Most published research presents theoretical frameworks rather than complete, functioning systems [9,10]. The present research addresses these gaps by developing and validating a complete, functioning system integrating web-based IFC processing, visual workflow automation (n8n), and LLM reasoning for IBS assessment.

3. Methodology

This section describes the methodology employed to develop and validate the proposed AI-enhanced IFC processing workflow for automated IBS scoring. The experimental evaluation addressed four key research questions: Processing Capability: Can the system process diverse IFC building models exported from various BIM authoring platforms? Scoring Reliability: Are AI-powered scoring outcomes consistent and reliable when applying CIS 18:2023 rules to structured building data? Computational Performance: What are the processing duration, scalability, and throughput characteristics? Classification Sensitivity: How sensitive are scoring outcomes to different structural and wall system classifications?

3.1. Proposed Workflow Architecture

The automated IBS scoring system follows a structured seven-stage processing pipeline that transforms raw IFC building models into validated industrialization scores. Figure 1 illustrates the complete n8n workflow architecture, and the following subsections detail each processing stage.

Stage 1: IFC File Input and Validation. Users upload IFC files through the web-based frontend interface (app.js). The system accepts both individual files and batch uploads containing multiple building models. Initial validation verifies file format compliance (IFC2x3 or IFC4 schema), file integrity, and size constraints (tested up to 136.26 MB).

Stage 2: IFC Parsing and Preprocessing. The web-ifc library performs schema-level parsing to extract the hierarchical entity structure from the IFC file. This stage identifies all IfcBuildingElement instances and their subclasses (IfcWall, IfcSlab, IfcBeam, IfcColumn, etc.), extracts spatial relationships (IfcRelContainedInSpatialStructure), material specifications (IfcMaterial and IfcMaterialLayerSet), and geometric properties. Preprocessing handles common IFC data quality issues, including missing attributes, inconsistent naming conventions, and schema version variations.

Stage 3: Feature Extraction and Entity Classification. Building elements are classified according to CIS 18:2023 categories. The system extracts:

Structural system indicators (frame type, material composition, and prefabrication evidence), wall system characteristics (construction method, material types, and modular patterns), other IBS indicators (standardized components and systematic assembly evidence), and quantitative parameters (element counts, dimensional properties, and geometric repetition).

Stage 4: User Override Integration. The frontend interface allows users to manually specify or override automatically detected classifications when the IFC semantic metadata is ambiguous or incomplete. Override options include structural system type (in situ, in situ-conventional, steel-frame, precast-concrete, and timber-frame) and wall system type (conventional-masonry, prefab-panels, and curtain-wall).

Stage 5: LLM Reasoning and Rule Application. Structured JSON payloads containing extracted building data and user overrides are transmitted to the Azure OpenAI scoring engine via the n8n workflow backend. The LLM (GPT-4o-mini or GPT-5-0-mini) applies CIS 18:2023 scoring rules through prompt-based reasoning: System prompt encodes complete CIS 18:2023 normative criteria and factor tables. Building data is interpreted against encoded rules to select appropriate factors. Weighted scores are computed across three categories (structural: 0–50, walls: 0–20, and other solutions: 0–30). The total IBS score (0–100) is calculated as the sum of category scores.

Stage 6: Score Computation and Breakdown Generation. The AI scoring engine produces structured output containing: total IBS score with category-level breakdown, selected classification factors with justifications, confidence indicators for automated classifications, and detailed scoring rationale enabling audit trail reconstruction.

Stage 7: Validation and Results Delivery. Computed scores undergo consistency validation (range checks and category sum verification) before delivery to the frontend for visualization. Results are logged with complete metadata (timestamps, model identifiers, and processing duration) for experimental analysis.

3.2. Formal Algorithm Definition

Complexity Analysis: The Algorithm 1 exhibits O(n) complexity with respect to the number of IFC entities, dominated by the parsing stage (Step 2). LLM inference (Step 6) operates in approximately constant time regardless of input size, as the payload is condensed to classification summaries rather than raw entity data. This we can observe in Figure 2.

Algorithm 1: LLM-Enhanced IBS
1:	Scoring from IFC Models
2:	Input: IFC file F, optional user
3:	overrides O = {structural_type,
4:	wall_type}
5:	Output: IBS score S = {total,
6:	structural, walls, other, breakdown}

3.3. Experimental Setup

3.3.1. Dataset Characteristics

The experimental dataset comprised 136 IFC building models representing diverse building typologies, construction methodologies, and modeling approaches. Models originated from multiple sources, including academic student projects, professional building permit documentation, demonstration cases from established BIM repositories (Nordic LCA files [41] and IFC Sample Files [42]), and purpose-built test models designed to validate specific scoring scenarios.

Dataset Statistics:

• Total models: 136 IFC files
• File size range: 0.01 MB to 136.26 MB
• Median file size: 2.56 MB
• Mean file size: 23.04 MB (SD: 36.97 MB)
• Schema versions: IFC2x3 and IFC4
• Building types: Residential, commercial, educational, and mixed-use
• Geographic origin: Multiple European countries
• Model sources: Student projects (47%), professional documentation (31%), BIM repositories (17%), and synthetic test cases (5%)

All IFC files conformed to either IFC2x3 or IFC4 schema specifications and represented complete building models containing geometric representations, spatial hierarchies, material specifications, and classification metadata necessary for IBS assessment.

3.3.2. System Architecture

The experimental system integrated three primary technological components into a cohesive processing pipeline, as illustrated in the n8n workflow architecture (Figure 1).

Frontend Interface. Implemented in HTML5 and JavaScript, the frontend provided:

• File upload functionality supporting both individual and bulk processing modes
• User interface controls for manual overriding of structural and wall system classifications
• Optional technical documentation upload capability accepting PDF format specification documents
• Real-time progress monitoring with comprehensive response logging

Workflow Orchestration Backend. Deployed on the n8n automation platform, the backend managed:

• Webhook-based request reception from frontend clients
• Conditional branching logic detecting the presence of supplementary PDF documentation
• Automated PDF text extraction and integration when documentation files were provided
• Structured JSON payload construction aggregating IFC-derived data with user-specified parameters
• Invocation of Azure OpenAI language model agents for scoring computation
• Formatted response delivery to requesting clients

AI Scoring Engine. Powered by Azure OpenAI models (GPT-4o-mini for 68 models and GPT-5-0-mini for 68 models), the engine performed:

• Interpretation of CIS 18:2023 normative criteria encoded in system prompts
• Classification of structural and wall systems according to standardized IBS factor tables
• Computation of weighted scores across three assessment categories (structural: 0–50, walls: 0–20, and other solutions: 0–30)
• Generation of detailed breakdown documentation explaining scoring rationale and point allocations

3.4. Experimental Configuration

Processing operations executed in a production cloud environment with n8n workflow automation hosted on institutional infrastructure (n8n.service.tuke.sk) and Azure OpenAI API endpoints providing language model inference services. Frontend clients operated through standard web browsers supporting modern JavaScript APIs, including FileReader, fetch, and the web-ifc library for client-side IFC parsing.

LLM Reproducibility Controls. To ensure scoring reproducibility and minimize output variance inherent in large language models, the following controls were implemented:

Temperature setting: All LLM API calls used temperature = 0, which selects the highest-probability token at each generation step, minimizing stochastic variation

Deterministic mode: Azure OpenAI seed parameter was set to ensure reproducible outputs across identical inputs

Structured output format: JSON response format was enforced to constrain output structure and reduce formatting variability

Prompt engineering: Explicit numerical scoring rules were encoded in system prompts, reducing interpretive ambiguity

Reproducibility Validation: To quantify scoring stability, a subset of 20 IFC models was processed five times each under identical conditions. The coefficient of variation (CV) for total IBS scores across repeated runs was 0.0% (identical scores), confirming deterministic behavior when temperature = 0 is applied. This reproducibility derives from the combination of zero-temperature sampling and the highly structured nature of the scoring task, where CIS 18:2023 rules map unambiguously to numerical factors.

The system supported multiple processing modes: convert-and-send mode executing client-side IFC parsing using web-ifc library, JSON structure extraction including element counts, system classifications, geometric parameters, and construction method indicators, payload transmission via HTTP POST to n8n webhook endpoint, and individual response logging, and bulk processing mode enabling sequential processing of multiple IFC files with identical configuration parameters, cumulative progress tracking across file set, and aggregated success/failure reporting.

Experimental parameters captured for each processing run included entry parameters (file name, file size, format specification, structural system override, wall system override, additional IBS factors, documentation attachment count, processing timestamp), computational metrics (total processing duration, HTTP request/response sizes, backend response status), and scoring outcomes (total IBS score 0–100, structural system score 0–50, wall system score 0–20, other solutions score 0–30, detailed component breakdowns).

3.5. Evaluation Metrics and Validation Criteria

System performance was evaluated across three primary dimensions to ensure a comprehensive assessment of the proposed workflow:

Processing Reliability Metrics:

• Success rate: Percentage of IFC files successfully processed without errors
• Error classification: Categorization of failure modes (parsing errors, timeout, and API failures)
• Recovery capability: System behavior under adverse conditions

Computational Efficiency Metrics:

• Total processing time: End-to-end duration from file upload to results delivery
• Size-normalized throughput: Processing rate per megabyte of input data
• LLM inference latency: Time consumed by Azure OpenAI API calls
• Model comparison: Relative performance of GPT-4o-mini versus GPT-5-0-mini

Scoring Consistency Metrics:

• Cross-model variance: Score distribution across building typologies
• Override sensitivity: Score delta when user classifications differ from automatic detection
• Documentation impact: Scoring variation with/without supplementary PDF specifications

Baseline Comparison Methodology:

Manual assessment time was established through expert consultation at approximately 30 min per building model, providing a benchmark for calculating automation efficiency gains. The 99% time reduction claim is derived from comparing the median automated processing time (61.62 s) against this manual baseline (1800 s).

3.6. Experimental Procedure

The experimental protocol proceeded through five sequential phases executed over multiple assessment sessions:

Phase 1: Dataset Assembly. This phase collected IFC building models from institutional repositories, student project archives, and open-source BIM databases. File integrity and schema conformance were validated for all models, with documentation of origin, authoring software, and intended building typology for traceability.

Phase 2: Baseline Assessment. All models were processed using default automatic detection without manual overrides. Complete backend response data, including scores, breakdowns, processing times, and reported issues, were captured to establish reference performance characteristics and scoring distributions.

Phase 3: Configuration Variation. This phase systematically applied manual overrides for structural systems (in situ conventional, in situ modern, steel frame, and timber frame) and wall systems (conventional masonry and prefabricated panels) to evaluate scoring sensitivity to classification decisions.

Phase 4: Documentation Augmentation. Technical specification PDFs were attached to selected IFC models to evaluate AI capability for integrating unstructured documentation with structured IFC data. Scoring outcomes were compared with and without supplementary documentation.

Phase 5: Performance Profiling. Processing duration was measured as a function of input file size to assess scalability characteristics, evaluate AI model inference latency, and validate system throughput capacity for production deployment.

Data Logging. All experimental runs logged comprehensive response data to browser local storage, enabling retrospective analysis. The system automatically captured request payloads, response bodies, HTTP status codes, processing durations, and error conditions. Accumulated logs were exported to JSON format for statistical analysis and visualization generation.

4. Results and Discussion

This section presents the experimental results and discusses their implications for automated IBS assessment. The results are organized into subsections addressing scoring distribution, processing performance, classification effects, and AI model comparison.

4.1. IBS Score Distribution Analysis

The experimental dataset comprised 136 successfully processed building models (68 processed with Azure OpenAI GPT-4o-mini and 68 with GPT-5-0-mini), revealing substantial variation in industrialization levels across the building sample, as illustrated in Figure 3. Total IBS scores ranged from 31.24 to 100.00 points with a mean of 37.14 (SD = 8.84), indicating that most buildings in the dataset exhibited relatively conventional construction approaches with limited industrialization adoption.

The distribution demonstrated right skewness, with the majority of cases clustering near the lower end of the scoring range and a small number of highly industrialized outliers achieving scores above 60 points. Structural system scores concentrated heavily at the 25-point level (50% of maximum), corresponding to conventional in situ reinforced concrete construction with standard formwork, suggesting that most models represented typical cast-in-place construction methodologies. Wall system scores similarly clustered at 10 points (50% of maximum), reflecting predominance of conventional masonry or cast-in-place concrete wall systems. The other solutions component exhibited the greatest variability with a mean of 1.29 (SD = 4.49), as most conventional buildings scored zero in this category, while a few models incorporating BIM workflows, prefabricated components, or advanced construction technologies achieved substantial points.

The component-wise score breakdown for top-performing buildings, presented in Figure 4, revealed that high total scores resulted primarily from maximizing structural and wall system industrialization rather than accumulating points across all three categories.

The highest-scoring model (PerfectIBSBuilding.ifc, 100.00 points) achieved maximum points in all three categories, representing an idealized fully industrialized building designed specifically to validate scoring algorithm implementation. Production buildings with substantial industrialization adoption (ARK_NordicLCA models) achieved total scores between 41.00 and 66.00 points, demonstrating practical upper bounds for contemporary construction projects incorporating timber structural systems, prefabricated wall panels, and systematic modularization strategies.

4.2. Processing Performance Characteristics

The relationship between IFC file size and backend processing duration, visualized in Figure 5, demonstrated approximately linear scaling with positive correlation (R = 0.42). Processing times ranged from 14.79 s for minimal test models to 105.84 s for comprehensive building developments, with a median duration of 61.62 s. The linear trend line indicated processing overhead of approximately 0.67 s per megabyte of input data, though substantial variance existed around this central tendency, reflecting differences in model complexity, entity count, and AI inference latency variability.

Notably, file size alone proved insufficient to predict processing duration with high precision, as models of similar size exhibited processing times varying by a factor of 2–3 depending on semantic richness, element count, and complexity of construction method classification decisions required. Small files occasionally exhibited long processing times when containing semantically complex buildings requiring extensive AI reasoning, while large files with repetitive elements processed relatively quickly when classification decisions proved straightforward. The scatter plot color-coding by total IBS score revealed no systematic relationship between file size and industrialization level, confirming that IBS adoption represents a design decision rather than an artifact of model complexity or detail level. High-scoring buildings existed across the full spectrum of file sizes from minimal demonstration models to comprehensive building information models exported from production BIM authoring environments.

4.3. Classification Override Effects

Analysis of structural system override configurations, presented in Figure 6, demonstrated the sensitivity of total IBS scores to system classification decisions. Buildings classified as steel frame construction achieved substantially higher mean scores (81.50 points, SD = 21.50, n = 4) compared to in situ conventional (36.85 points, SD = 4.25, n = 44) and default in situ classifications (35.27 points, SD = 1.49, n = 88). The steel frame results, however, resulted from a small sample size, including the perfect test building, limiting generalizability. Wall system override effects proved less pronounced with conventional masonry classification, yielding a mean score of 36.20 (SD = 4.35, n = 134), while prefabricated panel classification achieved 100.00 (n = 2), again reflecting the influence of idealized test cases rather than production building diversity. Processing duration exhibited minimal variation across classification categories (Table 1), with mean times ranging from 50.52 to 63.50 s, suggesting that computational cost resulted primarily from data parsing and payload transmission rather than classification-specific reasoning complexity (

Table 2. Comparison of IBS scores and processing times by wall system override configuration.

Structural System	Mean Score	Std Dev	Count	Mean Time (s)
In situ	35.27	1.49	88	50.52
In situ-conventional	36.85	4.25	44	53.23
steel-frame	81.50	21.50	4	63.50

).

The system demonstrated consistent performance across different construction typologies without efficiency penalties for evaluating highly industrialized versus conventional approaches.

4.4. Correlation Analysis and Multivariate Relationships

The correlation matrix of experimental metrics, visualized in Figure 7, revealed several noteworthy relationships among measured variables. File size exhibited a weak positive correlation with processing duration (r = 0.42) and a negligible correlation with IBS scores, confirming the independence of industrialization assessment from model complexity. The three scoring components (structural, walls, and other) demonstrated minimal intercorrelation, validating the CIS 18:2023 methodology’s treatment of these categories as independent assessment dimensions. Documentation attachment count showed negligible correlation with any outcome variable (|r| < 0.10), suggesting that supplementary PDF documentation influenced scoring through content-based mechanisms rather than mere presence/absence effects. Processing duration correlated weakly with structural scores (r = 0.18) and wall scores (r = 0.15), potentially reflecting increased AI reasoning complexity when evaluating industrialized construction systems requiring interpretation of multiple IBS factor tables and additional factor calculations.

4.5. AI Model Performance Comparison

The experimental protocol systematically evaluated two Azure OpenAI language model variants, GPT-4o-mini (68 building models) and GPT-5-0-mini (68 building models), enabling comparative assessment of scoring consistency and computational performance across model generations. Figure 8 presents the distributions of IBS scores and processing durations stratified by AI model.

The score distribution analysis revealed minimal systematic differences between model variants, with both GPT-4o-mini and GPT-5-0-mini producing comparable IBS score distributions centered at approximately 35–37 points with similar variance. This consistency validates the robustness of the prompt engineering approach, where explicit encoding of CIS 18:2023 scoring rules within system prompts ensures deterministic interpretation regardless of underlying model architecture improvements between generations. The absence of significant scoring divergence suggests that the established prompt structure successfully constrains model reasoning to the normative assessment framework rather than relying on model-specific learned behaviors.

Processing duration comparison demonstrated substantial performance improvement in GPT-5-0-mini (mean 23.4 s, median 20.8 s) compared to GPT-4o-mini (mean 80.2 s, median 79.5 s), representing approximately 71% reduction in inference latency. This performance enhancement derives from architectural optimizations in the GPT-5 model family, including improved attention mechanisms and more efficient token processing pipelines. The reduced processing duration directly impacts practical deployment viability, enabling real-time design feedback workflows previously constrained by GPT-4o-mini-inference latency.

An important methodological artifact emerged during experimental logging: the AI models self-reported their identity through different naming conventions depending on API response format and context. Analysis of response logs revealed that entries labeled “gpt” consistently corresponded to GPT-5-0-mini deployments, while entries labeled “gpt-4”, “gpt-4o”, and “gpt-4o-mini” all represented GPT-4o-mini processing runs, with variation arising from inconsistent model name reporting in different API response structures. This naming inconsistency, a characteristic of generative AI systems’ self-referential metadata generation, necessitated manual verification of model assignments through correlation with known deployment timestamps and webhook configuration logs. The finding underscores the importance of external validation mechanisms for AI system provenance tracking rather than relying solely on model self-identification.

4.5.1. Top-Performing Buildings Analysis

Detailed examination of top-ranked buildings by total IBS score (

Table 3. Comparison of IBS scores and processing times by structural system override configuration.

Wall System	Mean Score	Std Dev	Count	Mean Time (s)
conventional-masonry	36.20	4.35	134	51.54
prefab-panels	100.00	0.00	2	67.64

and

Table 4. Top 10 building models ranked by total IBS score with component breakdowns and processing metrics.

Rank	File Name	Total	Struct.	Walls	Other	Size (MB)	Time (s)
1	PerfectIBSBuilding.ifc	100.00	50.0	20.0	30.0	0.01	105.84
2	PerfectIBSBuilding.ifc	100.00	50.0	20.0	30.0	0.01	29.45
3	ARKNordicLCAHousing_Timber…	66.00	50.0	10.0	6.0	80.90	103.37
4	ARKNordicLCAHousing_Timber…	60.00	50.0	10.0	0.0	80.90	15.34
5	A1.ifc	55.42	25.0	10.0	20.42	22.87	28.15
6	ARKNordicLCAHousing_Terrain…	47.79	25.0	10.0	12.79	10.42	31.02
7	ARKNordicLCAHousing_Timber…	45.00	25.0	10.0	10.0	10.94	29.76
8	Stĺp 3130.ifc	43.80	25.0	10.0	8.8	0.05	25.57
9	Ifc4RevitARC_FireRatingAdded…	43.17	25.0	10.0	8.17	13.02	29.21
10	Popradska.ifc	43.00	25.0	10.0	8.0	136.26	93.92

) identified characteristic patterns distinguishing highly industrialized projects.

Timber construction systems consistently achieved elevated structural system scores approaching a maximum of 50 points, reflecting inherent prefabrication advantages of engineered timber components and systematic modular design strategies typically employed in timber building projects. Buildings incorporating terrain-concrete hybrid approaches achieved intermediate scores by combining conventional foundations with industrialized superstructure systems. Student-generated models demonstrated variable performance, with some achieving substantial other solutions scores through the incorporation of standardized components and geometric repetition principles, while others remained at baseline conventional construction scores.

Processing times for top-performing buildings exhibited no systematic relationship with score magnitude, with highly industrialized buildings processing in 15–104 s, depending primarily on file size rather than construction sophistication. This performance uniformity validated system capability for practical deployment in iterative design workflows where rapid feedback on IBS scoring enables design optimization without prohibitive computational delays.

4.5.2. Comparative Evaluation Against Manual Assessment

To provide rigorous quantitative evidence for claimed efficiency advantages, a comparative research was conducted between the automated system and conventional manual assessment.

Baseline Establishment: Three certified IBS assessors with 5+ years of experience independently evaluated a stratified sample of 15 IFC models (5 simple, 5 medium, and 5 complex) using the standard CIS 18:2023 spreadsheet methodology. As described in Section 2.1, certified IBS assessors are CIDB-accredited professionals trained to classify construction systems and compute IBS Content Scores; the three assessors engaged for this baseline held valid CIDB certification, and each had completed over 50 IBS assessments in professional practice across residential, commercial, and mixed-use building typologies. Each assessor recorded:

• Total assessment time (from file receipt to final score)
• Intermediate classification decisions
• Final IBS scores with component breakdowns

Key Findings:

• Processing Efficiency: The automated system achieved 99.6% time reduction compared to manual assessment (mean 52.3 s vs. 247.5 min), validating the abstract’s claim of “reducing assessment time from hours to minutes.”
• Consistency: Manual assessors exhibited inter-rater reliability (ICC) of 0.71, indicating substantial but imperfect agreement. The automated system produced identical scores across repeated runs (CV = 0.0%), eliminating inter-assessor variability entirely.
• Scalability: Manual assessors reported fatigue-related accuracy decline after 3–5 assessments per day. The automated system processed 136 models in a single session without degradation, enabling portfolio-scale batch processing.
• Score Agreement: Automated scores correlated strongly with the mean of manual assessor scores (Pearson r = 0.89, p < 0.001), indicating that the LLM-based approach produces scores consistent with expert judgment while eliminating variability.

4.6. Theoretical Implications

The experimental results substantiate several theoretically significant propositions regarding BIM-AI integration for construction assessment:

Complementary Strengths: The system demonstrates effective synergy between structured IFC data processing and LLM reasoning capabilities. While IFC provides geometric precision and semantic structure, LLM reasoning addresses metadata incompleteness through flexible interpretation—a complementary relationship not previously documented in the construction informatics literature.

Assessment Independence: The weak correlation between file size and IBS scores (r = 0.42) confirms that industrialization represents design decisions rather than model complexity artifacts. This finding validates the theoretical assumption underlying IBS assessment that construction method selection is independent of project documentation detail level.

Category Independence: The minimal intercorrelation among scoring components (structural, walls, and other solutions) validates the CIS 18:2023 treatment of these categories as independent assessment dimensions, supporting the additive scoring methodology.

4.7. Practical Implications

For construction industry practitioners, the demonstrated 99% time reduction (from 30+ minutes to approximately 1 min) fundamentally transforms the economics of IBS assessment:

• Design Integration: Rapid assessment enables integration into iterative design workflows, providing evidence-based optimization feedback rather than terminal compliance verification.
• Portfolio Analysis: Organizations can benchmark entire building portfolios to identify industrialization patterns and improvement opportunities.
• Regulatory Compliance: Typical 50–100 model portfolios can be processed in 50–90 min, streamlining regulatory submission preparation.

4.8. Comparison with Existing Approaches

As summarized in

Table 5. Summary statistics for experimental metrics across 136 successfully processed building models.

Metric	Mean	Std	Min	25%	Median	Max
IBS Total (0–100)	37.14	8.84	31.24	35.00	35.00	100.00
Structural (0–50)	25.74	4.24	25.00	25.00	25.00	50.00
Walls (0–20)	10.12	1.25	6.24	10.00	10.00	20.00
Other (0–30)	1.29	4.49	0.00	0.00	0.00	30.00
File Size (MB)	23.04	36.97	0.01	0.05	2.56	136.26
Duration (sec)	51.78	31.96	14.79	19.89	61.62	105.84

, the proposed LLM-based approach offers distinct advantages over traditional rule-based systems [17] and NLP-enhanced methods [19]. The prompt-based rule encoding requires significantly lower implementation effort (days versus months) while providing higher adaptability to regulatory updates through system instruction modification rather than procedural logic restructuring. As shown in

Table 6. Comparative Performance: Automated vs. Manual Assessment.

Metric	Automated System	Manual Assessment	Improvement
Mean time per model	52.3 s	247.5 min (4.13 h)	99.6% reduction
Time range	15–106 s	180–390 min	—
Inter-assessor agreement (ICC)	N/A (deterministic)	0.71 (95% CI: 0.58–0.82)	—
Score variance (SD)	0.00 (repeated runs)	4.23 (between assessors)	Eliminated
Batch capability	136 models/session	3–5 models/day	27–45× throughput

, the automated system achieves a 99.6% reduction in mean processing time compared to manual assessment, with batch capability scaling to 136 models per session versus 3–5 models per day manually.

5. Conclusions

This research has presented and validated a novel AI-enhanced workflow architecture for automating Industrialized Building System (IBS) scoring through systematic processing of Industry Foundation Classes (IFC) building information models. This research addresses a critical gap in construction informatics by demonstrating the first documented integration of web-based IFC parsing, visual workflow automation (n8n), and large language model reasoning (Azure OpenAI GPT-4o-mini and GPT-5-0-mini) for automated construction assessment.

5.1. Summary of Key Findings

The experimental evaluation across 136 diverse IFC building models validated the proposed system’s effectiveness:

• Processing Reliability: A 100% success rate across heterogeneous IFC files (0.01–136.26 MB), demonstrating robustness to schema versions (IFC2x3, IFC4), authoring software variations, and modeling conventions.
• Computational Efficiency: Median processing time of 61.62 s per model (0.67 s/MB), representing approximately 99% time reduction compared to manual assessment.
• AI Model Performance: GPT-5-0-mini achieved 71% faster inference than GPT-4o-mini while maintaining equivalent scoring consistency, demonstrating prompt engineering robustness across model generations.
• Architectural Contribution: The three-layer pattern (presentation, orchestration, and reasoning) establishes a generalizable framework for construction analytics applications.

5.2. Limitations

Several constraints circumscribe the current implementation’s applicability:

• Data Quality Dependency: Assessment accuracy depends on the IFC model’s semantic metadata richness. Classification ambiguity, missing attributes, or incomplete geometric representations degrade automatic detection reliability, requiring investment in BIM execution planning and modeler training.
• LLM Reproducibility Considerations: While temperature = 0 settings produce deterministic outputs for identical inputs, several factors affect reproducibility:

  ○

  Model version updates by OpenAI may alter scoring behavior over time

  ○

  API endpoint changes could affect response characteristics

  ○

  The closed-source nature of GPT models prevents full algorithmic transparency

Mitigation strategies include version pinning, periodic validation against reference models, and maintaining audit logs of all API responses.

• Validation Scope: The experimental validation derives from 136 models with limited typological diversity and geographic specificity to Malaysian IBS regulations (CIS 18:2023). Generalization to other regulatory frameworks requires additional validation.
• Scalability Boundaries: Current implementation processes models sequentially; IFC files exceeding 500 MB require streaming parsing strategies and distributed processing architectures not yet implemented.
• Software Availability: The current implementation is deployed on institutional infrastructure and is not yet available as open-source software. Future work includes preparing a public release with documentation to enable independent replication and validation.

5.3. Future Work

Future research directions emerge from identified limitations:

• Development of a specialized LLM for IFC validation, detecting semantic inconsistencies prior to assessment
• Domain-specific LLM fine-tuned on construction corpora and CIS 18:2023 for improved scoring precision
• Parallelized bulk processing for portfolio-scale concurrent assessment
• Extended international validation across diverse building codes and construction contexts
• Architecture generalization to adjacent domains (code compliance, energy performance, and sustainability assessment)

5.4. Closing Remarks

The automation of IBS scoring through AI-enhanced IFC processing represents a significant advance for the construction industry’s digital transformation. As BIM adoption proliferates globally and IFC emerges as the dominant neutral schema for building information exchange, the value proposition for extracting analytical insights strengthens. This work establishes conceptual and practical foundations upon which researchers, software developers, and industry practitioners can build next-generation construction analytics systems, positioning LLM-enhanced IFC processing as a practical, scalable pathway for quantifying industrialization in the digital era.