Large Language Models (LLMs) are emerging as natural-language interfaces for Building Information Modeling (BIM), successfully achieving information retrieval, but no benchmark exists to evaluate their ability to edit Industry Foundation Classes (IFC). We introduce BIBIMBAP, a Text-to-BIM benchmark with 100 curated tasks covering Create-Read-Update-Delete (CRUD) operations across spatial, geometric, topological, numeric, and conceptual categories. Each task includes a natural-language prompt, an IFC model, expected structured outputs, and test scripts for automated evaluation. Baseline results show limited performance (best: 50.2%) and frequent failures in constraint preservation and spatial reasoning, with some correct outputs achieved despite flawed reasoning.
Current building energy simulations often rely on static assumptions, neglecting dynamic factors such as climate change and component degradation, leading to performance gaps and suboptimal retrofit decisions. This research applies a multi-modelling approach that integrates climate projections with building envelope and HVAC degradation models within an automated workflow. Using an office building case study, the results demonstrate the efficacy of multi-modelling while also improving understanding of maintenance–retrofit trade-offs under climate uncertainty. The analysis reveals key sensitivities and limitations of dynamic energy modelling, advancing insight into how multi-modelling can support climate-aware, lifecycle-oriented retrofit decision-making.
BIM-based model checking helps improve the quality of data and information throughout the planning phase of a built asset. However, rule-based checking can be time-consuming for large, frequently updated BIM models. Increment descriptors for BIM models are enriched with
engineering intent, thereby unveiling additional information about the changes made. The proposed method enhances the efficiency of model checking by executing only the rules affected by changes, rather than reprocessing the entire model. Overall, this results in better traceability of issues and edits, suggesting future opportunities to align change actions with the deficiencies they address.
Bridge management is increasingly automated through continuous monitoring, AI-based diagnostics, and digitally coordinated interventions. Yet, current workflows rarely provide end-to-end verifiability: sensor datasets, off-chain computations, and resulting maintenance decisions are difficult to audit across multiple stakeholders. This paper proposes a trust-layered architecture that combines decentralized storage, verifiable computation, and blockchain governance to make monitoring-to-action pipelines tamper-evident and reproducible. The paper discusses how smart contracts can orchestrate inspection and maintenance playbooks, while computation integrity can be certified via Trusted Execution Environments or zero-knowledge proofs. Trade-offs in scalability, privacy, and compliance are analyzed, outlining practical deployment patterns for bridge lifecycles.
Manual regulation compliance checking remains common in architectural design, while BIM-based and semantic approaches operate outside the authoring environment, requiring specialist expertise, delaying feedback in iterative work. This paper presents a lightweight compliance-checking framework implemented as a pyRevit extension and prototyped in Autodesk Revit. The framework extracts model parameters, converts them to RDF aligned with ifcOWL and a project namespace, and validates constraints using Shapes Constraint Language. The case study shows non-compliances can be detected and reported via both documentation outputs and in-model graphical feedback. The contribution is an auditable linked-data mechanism for extensible regulation checking within BIM authoring tools.
Creating synthetic datasets is challenging as they need to represent the diversity of real-world sites and measured point clouds. This paper introduces SYN3D-LLM, a dialog-driven large language model framework for generating semantically annotated synthetic 3D point clouds. By producing executable code through an iterative designer-critic loop. Experimental evaluations of the framework demonstrate that it improves the pass rate by discussing and testing generated code and by assessing the geometrical consistency between the initial prompt and the generated data. It also demonstrates that SYN3D-LLM is versatile in generating point clouds for various cases.
Existing generative design approaches for floorplan generation rarely incorporate fabrication constraints, limiting their applicability in industrialized construction. This study presents a fabrication-aware generative design framework for hybrid prefabrication based on Stable Diffusion XL. The framework integrates spatial, functional, and manufacturing constraints to balance design flexibility and production efficiency. A two-stage training strategy generates both conceptual layouts and module-integrated solutions. Experiments demonstrate the generation of boundary-compliant layouts and standardized module allocation under diverse boundary conditions and text-based requirements. An interactive workflow supports convergence from conceptual designs to fabrication-aligned solutions, bridging early-stage architectural intent with downstream fabrication requirements.
Existing Lean construction simulation games rarely provide detailed personalized feedback for participants but rely on coarse metrics like completion time or flow visualization. This paper investigates how high-resolution telemetry in active multi-role Virtual Learning Environment (VLE) supports lean assessment on teams in a staged introduction of process and automation improvements. We developed a VLE for 11 building trades using role-specific input devices and a telemetry pipeline that captures interactions to compute Lean indicators (flow discontinuity, production rate, waste). We found high potential of embedding high-resolution telemetry into VLE of data-driven evaluation of Lean principles, offering value to educators, trainers, and practitioners.
Bridge conceptual design is challenged by heterogeneous boundary conditions, strict engineering constraints, and large solution spaces. This paper presents an AI-assisted pipeline for automated bridge conceptual design, focusing on early-stage span layout planning. The proposed approach integrates ontology-based semantic knowledge representation with reinforcement learning to enable automated decision-making while preserving engineering consistency. By embedding engineering constraints into the learning environment, the method supports efficient exploration of feasible span configurations. The results indicate that the proposed pipeline provides a robust and extensible foundation for AI-supported bridge planning in realistic engineering scenarios.
Digital Twins increasingly rely on real-time IoT data, yet the end-to-end architectural implications of alternative integration approaches remain underexplored. This paper presents a controlled comparative analysis of three IoT data pipelines connecting The Things Network (TTN) to a Unity-based Digital Twin for mixed reality visualisation: cloud-centric pipelines using Amazon Web Services and Microsoft Azure, and a lightweight publish–subscribe pipeline based on direct MQTT delivery. By isolating the data integration layer, a benchmarking approach evaluates data transmission behaviour, persistence, integration complexity, scalability, and cost. The findings provide practical architectural insight into how pipeline design choices shape Digital Twin behaviour.
