PaperlayerView exact revision on arXiv

Source-linked AI summary

GPT Models in Construction Industry: Opportunities, Limitations, and a Use Case Validation

Abdullahi Saka, Ridwan Taiwo, Nurudeen Saka, Babatunde Salami, Saheed Ajayi, Kabiru Akande, Hadi Kazemi

arXiv:2305.18997v1cs.HCcs.AIcs.CL

TL;DR

The study addresses limited knowledge about GPT models’ opportunities and challenges in the construction industry. It combines critical review, expert discussion, and case-study validation, finding applications across the project lifecycle while identifying deployment limitations and validating a materials-selection and optimization use case.

  • Problem

    The study examines the limited knowledge about opportunities and challenges associated with applying GPT models in the construction industry.

  • Method

    The study uses critical review, expert discussion through a modified Delphi survey, and case-study validation to assess GPT models in the AEC industry.

  • Results

    GPT opportunities span pre-design, design, construction, and post-construction, while the study also identifies challenges and validates a materials-selection and optimization use case.

  • Takeaways & Limitations

    The findings provide researchers and stakeholders with identified application areas, deployment challenges, and a research agenda for GPT models in construction.

  • Takeaways & Limitations

    GPT models have limited domain-specific construction knowledge, while changing regulations and country-specific requirements constrain their application without adequate fine-tuning and contextual information.

Abstract

from arXiv · show

Large Language Models(LLMs) trained on large data sets came into prominence in 2018 after Google introduced BERT. Subsequently, different LLMs such as GPT models from OpenAI have been released. These models perform well on diverse tasks and have been gaining widespread applications in fields such as business and education. However, little is known about the opportunities and challenges of using LLMs in the construction industry. Thus, this study aims to assess GPT models in the construction industry. A critical review, expert discussion and case study validation are employed to achieve the study objectives. The findings revealed opportunities for GPT models throughout the project lifecycle. The challenges of leveraging GPT models are highlighted and a use case prototype is developed for materials selection and optimization. The findings of the study would be of benefit to researchers, practitioners and stakeholders, as it presents research vistas for LLMs in the construction industry.

Graphical Abstract

GPT models emerged from advances in large language models and are being considered for construction’s information-intensive, low-productivity environment. The study reviews their opportunities and limitations in the AEC industry and validates a materials-selection use case.

  • AEC construction faces slow innovation adoption, fragmented information management, and productivity growth of only 1% per year over two decades.The industry also faces delays, safety concerns, cost overruns, labor shortages, and stringent government requirements.
  • GPT models are large language models based on transformer architectures that learn statistical patterns from large datasets.Their data-driven approach differs from earlier rule-based NLP systems requiring explicit grammar and syntax programming.
  • The study addresses limited AEC research by identifying GPT opportunities, evaluating application limitations, and validating a use case.Its stated objectives cover opportunities, limitations, and case-study validation in the construction industry.
  • GPT opportunities span the project lifecycle, while the paper also discusses value-added services and a case study for material selection and optimization.The study presents these opportunities alongside deployment challenges for researchers, practitioners, and stakeholders.
  • GPT models support language generation, sentiment analysis, text categorization, and question answering, but may inherit bias and require costly computation.Fine-tuning enables adaptation to a range of tasks, while deployment and training can remain computationally difficult and expensive.

3.0 Methodology

The study uses a qualitative, sequential methodology combining literature exploration, expert discussion, critical review, and validation of a GPT-based materials-selection use case.

  • Research approach: The research follows sequential phases comprising literature review, critical review, expert discussion, and a case study.The approach is qualitative and is depicted in Figure 2.
  • Initial exploration: Phase I searches Scopus, ACM, Web of Science, Science Direct, and Google Scholar to examine GPT applications in the AEC industry.The search identified only two publications specifically leveraging GPT models at that stage.
  • Expert discussion and critical review: Phase II combines expert discussion with critical review to assess GPT opportunities and limitations in the AEC industry.The critical review evaluates expert discussion findings alongside relevant existing studies.
  • Expert discussion and critical review: The expert discussion uses a modified Delphi approach involving ten panellists with at least 10 years of AI and AEC experience.Experts were selected using predefined domain-expertise and experience criteria.
  • Use case: Phase III evaluates GPT for material selection and optimization through a proposed system architecture followed by verification and validation.Verification and validation are conducted using checklists and case testing.

4.2 Expert Discussion

Expert discussion identified GPT opportunities across the construction project lifecycle, alongside both inherent model limitations and industry-specific deployment challenges.

  • Project lifecycle opportunities: GPT opportunities span project lifecycle phases from design and construction through operation and maintenance, including facility support, sustainability, and asset management.The expert categorization includes occupant communication, regulatory compliance, performance monitoring, energy management, incident resolution, and lifecycle asset management.
  • Construction and delivery: Construction-phase applications include resource allocation, site safety management, change-order management, quality assurance, documentation, dispute resolution, and cost planning.The opportunities were categorized by experts across operational functions relevant to project delivery and control.
  • Operation and maintenance: Operation and maintenance opportunities include predictive maintenance, energy optimization, incident resolution, occupant support, and regulatory compliance management.These applications target facility performance, occupant requests, compliance evaluation, and timely handling of operational incidents.
  • Demolition and redevelopment: GPT applications also cover demolition and redevelopment through demolition protocols, regulatory permits, environmental impact analysis, and material recovery.The categorization identifies opportunities for managing late-lifecycle activities and environmental considerations.
  • Limitations and deployment requirements: The identified limitations include capital and infrastructure requirements, scalability, cybersecurity, interdisciplinary and cultural considerations, and latency.The study distinguishes limitations inherent to GPT models from industry-specific challenges, with application requirements ranging from zero-shot learning to fine-tuning and knowledge-source integration.

5.0 Opportunities

GPT opportunities span the construction project lifecycle and extend to value-added services. The paper organizes these opportunities by lifecycle phase.

  • Opportunities are categorized across predesign, design, construction, operation and maintenance, demolition, and value-added services.
  • The paper presents opportunities for leveraging GPT models within each identified construction lifecycle phase.
  • The following subsections explain the opportunities and how GPT models can be applied.

5.1 Pre-Design Phase

The pre-design phase provides opportunities for GPT models to support project framing, procurement, requirements development, and management planning. Proposed uses emphasize faster, more informed, and better-coordinated decisions across stakeholders.

  • 5.1 Pre-Design Phase: GPT models could enhance pre-design by providing guidance on design and building procedures while improving decision-making, communication, and identification of project constraints.The proposed benefits include reduced human error and bias, stronger alignment with project objectives, and faster discovery of design restrictions and possibilities.
  • 5.1 Pre-Design Phase: GPT models could support procurement decisions by evaluating project data and producing suggestions for more accurate, data-driven assistance.This addresses existing procurement approaches that depend heavily on subjective expert knowledge and are vulnerable to human mistakes.
  • 5.1 Pre-Design Phase: GPT models could help develop project briefs and Employer Information Requirements by supporting stakeholder engagement and information management.These activities establish client requirements and provide a foundation for the wider construction process.
  • 5.1 Pre-Design Phase: Fine-tuned GPT models could support scope definition, scheduling, resource allocation and estimation, risk management, and decision support during project management planning.Prompting can help interpret diverse project factors and provide recommendations for effective management.

5.2 Design Phase

The design phase offers GPT applications for generating alternatives, checking compliance, estimating quantities and costs, improving energy performance, and supporting project information and safety activities. These uses target time-consuming, error-prone, and information-intensive design and delivery tasks.

  • 5.2 Design Phase: GPT models could accelerate design-concept generation by helping produce alternatives that satisfy project constraints, reducing reliance on time-consuming and subjective expert development.The proposed use addresses the traditional dependence on architects’ and engineers’ expertise and experience.
  • 5.2.2 Automated Regulatory Compliance: GPT models could automate regulatory compliance checks by comparing architectural and structural designs with building codes and identifying issues before construction.The proposed process is intended to reduce errors, streamline design, save time, and minimize costly revisions.
  • 5.2 Design Phase: GPT models could support quantity take-off and costing by processing project designs, materials, specifications, cost databases, and estimation methods.The proposed application targets an activity commonly assigned to quantity surveyors.
  • 5.2 Design Phase: GPT models could support energy-efficient building design using standards, regulations, passive-design principles, facade optimization, and renewable-energy-system information.The application responds to construction’s rising energy demand and the building envelope’s influence on energy consumption and loss.
  • 5.2 Design Phase: GPT models could assist design specification, risk management, progress reporting, site safety, resource allocation, change management, quality control, and dispute-related activities.The proposed uses include document generation, data analysis, hazard identification, impact analysis, inspection planning, anomaly detection, and centralized information management.

5.4 Operation and Maintenance Phase

The operation and maintenance phase presents GPT opportunities for predictive maintenance, energy and asset management, incident handling, facility communication, compliance, space use, sustainability, waste management, and emergency response. These applications rely on analyzing operational data and producing timely recommendations or documentation.

  • 5.4 Operation and Maintenance Phase: GPT models could enable predictive maintenance by analyzing historical data, sensor readings, and related factors to identify indicators of potential equipment failures.This would allow maintenance teams to address issues proactively rather than relying only on scheduled inspections or equipment failures.
  • 5.4 Operation and Maintenance Phase: GPT models could analyze energy-consumption patterns, weather, occupancy, and equipment performance to identify energy-saving opportunities and optimize usage.The proposed application supports real-time monitoring and analysis of building energy consumption.
  • 5.4 Operation and Maintenance Phase: GPT models could analyze and classify incident reports, generate standardized documentation, and help stakeholders prioritize resources for faster resolution.The proposed use targets manual, paper-based, and labor-intensive incident reporting processes.
  • 5.4 Operation and Maintenance Phase: GPT models could streamline asset management by predicting remaining useful life and supporting proactive maintenance and repair scheduling.The intended benefit is to reduce the risk of unexpected failures while improving asset performance and operational efficiency.
  • 5.4 Operation and Maintenance Phase: GPT-powered facility chatbots could collect and sort occupant maintenance requests and general enquiries, enabling faster communication with facility-management personnel.The described prototype uses zero-shot learning to organize occupant requests.
  • 5.4 Operation and Maintenance Phase: GPT models could support facility compliance, space optimization, sustainability, waste management, recycling, and emergency response by analyzing operational data and regulatory or procedural knowledge.Proposed outputs include non-compliance reports, space-reconfiguration decisions, waste-sorting recommendations, and emergency-response actions.

5.5 Demolition Phase

The demolition phase presents GPT opportunities across planning, risk assessment, environmental management, compliance, structural analysis, and materials recovery. These applications target labor-intensive, error-prone tasks involving safety, sustainability, and resource use.

  • Demolition Phase: GPT models can analyze project plans, site conditions, and historical records to identify hazards, optimize demolition sequences, and recommend safety measures.Their proposed role includes risk prediction and assessment.
  • Demolition Phase: GPT models can support waste management by analyzing demolition materials and improving recycling, disposal, and recovery decisions.The opportunity addresses manual sorting and decision-making that can be inefficient and error-prone.
  • Demolition Phase: GPT models can assist redevelopment planning by integrating site conditions, market demands, regulatory requirements, and other constraints.This is presented as an alternative to resource-intensive manual research, analysis, and stakeholder consultation.
  • Demolition Phase: GPT models can interpret regulations and codes, identify required permits, and provide real-time guidance for demolition compliance.The proposed automation aims to reduce errors in analyzing regulatory requirements.
  • Demolition Phase: GPT models can enhance environmental assessments and structural analysis by processing diverse data to identify risks, hidden defects, degradation, or material weaknesses.Relevant inputs include environmental databases, scientific literature, historical project data, maintenance records, and sensor data.
  • Demolition Phase: GPT models can improve demolition risk assessments by analyzing project records, structural characteristics, accident data, and safety guidelines to identify high-risk areas and failure modes.The proposed approach is intended to make assessments more accurate and objective than reliance on expert judgment alone.

5.6 Value-added Services

GPT models are presented as value-added services for construction firms beyond a single project phase. Proposed uses include knowledge management, customer and stakeholder communication, market analysis, and conversational interfaces.

  • Value-added Services: GPT models can create construction knowledge bases from best practices, regulatory requirements, and organizational information.They can also capture and preserve tacit expertise from historical data and past project reports for newer professionals.
  • Value-added Services: GPT-powered chatbots can provide immediate customer assistance, answer project and pricing inquiries, and generate tailored responses based on customer intent.The proposed applications target customer experience and satisfaction.
  • Value-added Services: GPT models can support stakeholder management by generating project updates, executive summaries, feedback analyses, and multilingual communication.They may be integrated with communication platforms for personalized stakeholder interaction.
  • Value-added Services: GPT models can analyze market reports, news, online posts, and customer sentiment to identify changing market dynamics, demands, technologies, and trends.The stated use is to help organizations improve products and services and align with market developments.
  • Value-added Services: GPT models can support conversational systems through natural-language recognition, intent classification, entity extraction, generation, and BIM information retrieval.Compared with rigid pattern-based systems, they may reduce data and engineering requirements for prototypes.

6.0 Challenges

The paper classifies challenges to deploying LLMs in construction into industry-related, LLM-related, and combined industry–LLM categories. Figure 11 presents this three-part challenge structure.

  • Challenges: Construction LLM deployment challenges are grouped into industry-related, LLM-related, and industry–LLM nexus categories.This classification is presented in Figure 11.

6.1 Hallucinations

The paper identifies hallucination, data and interoperability constraints, limited domain knowledge, confidentiality concerns, and weak trust as barriers to construction deployment. These risks are consequential because construction decisions affect safety, cost, quality, schedule, and legal responsibilities.

  • Hallucinations: GPT hallucinations can produce plausible but false information, creating risks for safety management, robotics, scheduling, project cost, and delivery time.The paper gives a fabricated product example and notes that incorrect outputs could endanger people and property.
  • Data and Interoperability: Construction data needed for GPT fine-tuning may be unavailable, heterogeneous, or lost because projects have unique attributes and digital data collection remains limited.Data availability and quality are identified as major barriers to AI application in the industry.
  • Data and Interoperability: Interoperability requires costly conversion among construction formats, while converting BIM data to GPT-supported formats may discard information.The paper states that interoperability costs can reach millions of pounds and lists formats including PDF, CAD, IFC, JSON, and CSV.
  • Domain-specific knowledge and Regulatory Compliance: GPT models have limited construction-specific knowledge and require fine-tuning, contextual information, and updates as technical regulations change.Regulations are numerous, technically drafted, context-dependent, and sometimes pictorial, increasing interpretation demands.
  • Confidential and Intellectual property: Using project designs, costs, contracts, schedules, and patented techniques raises confidentiality, intellectual-property, data-ownership, ethical, and legal concerns.The paper calls for clear policies and guidelines governing GPT use in construction.
  • Trust and acceptability: Industry resistance, limited trust in black-box systems, data-release concerns, job fears, and perceived tool complexity can hinder GPT deployment.Acceptance challenges may involve multiple project parties and can also affect access to data for fine-tuning.

6.6 Liability and Ethics

GPT deployment in construction creates liability, ethical, skills, infrastructure, scalability, cybersecurity, collaboration, and accountability challenges. These risks require attention to data quality, transparency, training, security, and responsible implementation.

  • Liability and accountability: GPT outputs may be biased, incomplete, inaccurate, legally non-compliant, and difficult to interpret, complicating liability and accountability.Unclear model processes can reduce trust and make responsibility for harmful recommendations difficult to assign.
  • Ethics and social impact: GPT models can amplify training-data bias, enable unethical use, and affect construction employment, making diverse and unbiased training data important.The passage links these concerns to accuracy, reliability, and labour-market effects.
  • Skills and training: Effective deployment requires prompt engineering, data-preprocessing, fine-tuning, and training programs that construction professionals may not yet possess.Zero-shot, few-shot, and chain-of-thought prompting are identified as relevant techniques.
  • Infrastructure and cost: Computing, connectivity, storage, and token-based usage costs can burden small and medium-sized construction enterprises.The passage notes that these firms represent about 80% of industry organizations.
  • Scalability: Fine-tuned GPT models remain difficult to scale because use-case-specific training depends on structured data and additional computational resources.GPT models reduce some Conversational AI scalability challenges, but fine-tuned models retain scalability constraints.
  • Cybersecurity: Cybersecurity exposure increases through interconnected systems, cloud computing, and data exchange, requiring training, monitoring, and incident response.The identified risks include unauthorized access, data breaches, and cyber-attacks.
  • Interdisciplinary collaboration: GPT-generated technical jargon can create misunderstandings across architecture, engineering, construction management, and data-science disciplines.Specialized terminology may hinder interdisciplinary collaboration when professionals do not share the same vocabulary.

6.12 Cultural and social considerations

GPT adoption in construction is shaped by cultural, social, linguistic, maintenance, latency, and infrastructure considerations. These factors affect acceptance, communication, system performance, and operational reliability.

  • Cultural and social considerations: Cultural and social differences among stakeholders can influence the acceptance, adoption, and effectiveness of GPT-driven construction solutions.Construction projects operate across varied cultural and social contexts with distinct values, norms, and practices.
  • Social acceptance: Concerns about job displacement, data privacy, ownership, and security can create resistance and undermine trust in GPT interventions.These concerns may hinder broader acceptance of GPT-enabled systems.
  • Latency: Latency is problematic for time-sensitive construction projects because GPT processing can require extensive training data and costly high-performance hardware.Limited processing power can particularly affect smaller construction firms.
  • Maintenance: GPT systems require ongoing maintenance and updates because changing data patterns, user requirements, and industry standards can cause performance drift.Degradation may reduce accuracy and reliability over time.
  • Technical infrastructure: Managing servers, databases, networks, and data-processing systems becomes more difficult as GPT applications increase in scale and complexity.Dedicated IT personnel, health checks, troubleshooting, and prompt response are identified as mitigation measures.
  • Multilingual communication: Multilingual construction projects need language-processing capabilities, but models trained in limited languages may struggle with less-represented languages.Language variability can hinder seamless communication among geographically distributed stakeholders.
  • Multilingual evaluation: ChatGPT underperformed task-specific models in various languages except English across summarization, question answering, named entity recognition, and part-of-speech tagging.The cited experiments used zero-shot learning to examine multilingual capacity.
  • Use-case architecture: The prototype’s architecture uses a BIM model, Forge services, and the Model Derivative API to extract searchable geometric and metadata information.The BIM file is uploaded to Forge, translated into SVF2, and rendered through the Forge Viewer.

7.2 NLP Prompt Processing Module

The prototype combines BIM-derived element properties, prompt engineering, ChatGPT, and an interactive web interface for material selection and optimization. Iterative prompting supported zero-shot, few-shot, and edge-case testing.

  • NLP Prompt Processing Module: The prompt manager connects to ChatGPT 3.5 Turbo through AJAX and uses a temperature of 0.5 to reduce response randomness.Prompt development is described as a multistep process for material selection and optimization.
  • NLP Prompt Processing Module: The system prompt is combined with BIM-extracted element properties, while the user prompt states the material-selection intention.Providing sufficient context helps guide the interaction.
  • NLP Prompt Processing Module: Prompt instructions and dialogue design are iteratively refined through zero-shot and few-shot testing, performance evaluation, and user feedback.The stated goal is improved system efficiency and effectiveness over time.
  • User interface and integration Module: The interface uses JavaScript, HTML, CSS3, Bootstrap 5, AJAX, and Forge Viewer to connect user queries, ChatGPT responses, and BIM visualization.The web interface supports communication between the front end and prompt manager.
  • User interface and integration Module: The prototype supports cloud deployment and mobile access while integrating BIM data with OpenAI services.The interface is designed for high availability and accessibility across devices.
  • Discussion: The interface displays user queries and ChatGPT responses alongside a 3D Autodesk Forge representation of the BIM model.This combines conversational interaction with contextual visualization of building objects.
  • Prompting scenarios: The prototype was tested through zero-shot, few-shot system-prompting, and edge-case scenarios.The zero-shot test retained the user prompt and BIM information while disabling the system-role prompt.
  • Prompting scenarios: Few-shot prompting produced context-sensitive material recommendations, while edge-case prompting restricted unrelated questions and the prototype still required component identification.Examples included bathroom-door material selection, an insulated wall recommendation with an R-value up to 20, and refusal of an unrelated BIM question.

8.0 Conclusion

The study addresses limited knowledge of GPT opportunities and limitations in construction through a preliminary review, expert discussion, and validated BIM-integrated prototype. It identifies lifecycle-wide opportunities while emphasizing the need for further validation and research.

  • Conclusion: GPT applications remain relatively underexplored in construction despite reducing some barriers to developing Conversational AI systems.The study responds to limited knowledge about construction-sector opportunities and limitations.
  • Conclusion: The study uses a preliminary literature review, expert discussion, and a validated material-selection and optimization prototype integrating BIM and GPT.The three-step approach connects opportunity identification with a practical use case.
  • Conclusion: GPT opportunities span predesign, design, construction, post-construction, and value-added services, using prompting or fine-tuning depending on the use case.The conclusion identifies zero-shot, few-shot, and chain-of-thought prompting alongside structured-data fine-tuning.
  • Conclusion: The study’s limitations include database and English-language search boundaries, a limited expert discussion sample, and no quantitative validation of the prototype.The authors identify prompt development, fine-tuning, database integration, and further opportunity exploration as future research directions.
Loading 2305.18997v1…