AI Architecture Research: Floor Plans, BIM & Building Performance
Read 116 author-sourced research records through specific architectural questions, dates and evaluation limits.
These records summarize author-reported research, including preprints and supporting datasets. They do not establish available product features or independently reproduced performance. Publication and revision dates are distinct.
116 records · Page 1 / 5
Text-Aided Multi-Modal Panoptic Symbol Spotting for CAD Floor Plan Drawings
RESEARCH /
The study uses both drawing primitives and annotations to identify floor-plan symbols. It addresses the gap between graphical form and the semantic information carried by text. A useful document-conversion test should inspect annotation hierarchy, object identity and conflicts between labels and geometry.
Architectural question
When do annotations change the meaning of a CAD symbol?
Mitigating Domain Shift in Conditioned Floor Plan Generation: Synthetic Pre-training for Data-Efficient Adaptation
RESEARCH /
The paper studies floor-plan generation across datasets with different regional and construction characteristics. It examines synthetic pretraining as an adaptation strategy. Architectural relevance lies in testing whether a model's learned habits transfer to a new brief without reproducing unsuitable room arrangements or construction assumptions.
Architectural question
Which layout conventions fail when a generator moves between regions?
BIM-Edit: Benchmarking Large Language Models for IFC-Based Building Information Modeling
RESEARCH /
BIM-Edit evaluates natural-language editing of existing IFC models rather than only creating new geometry. Its focus includes semantics and relationships. Architectural teams can use the benchmark question to assess whether an edit changes the intended elements while preserving identity, placement and unaffected model information.
Architectural question
What should an IFC edit preserve beyond visible geometry?
Bharathi Kannan Nithyanantham, Clemens Kujat, Tobias Sesterhenn, Stefan Telgmann, Ashwin Nedungadi, Jörn Plönnigs, Christian Bartelt, Stefan Lüdtke
PIRS: Physics-Informed Reward Shaping for SAC-Based Building Energy Management
RESEARCH /
PIRS places a PMV formulation inside a reinforcement-learning reward instead of relying only on temperature deviation. It studies how the objective shapes the energy-comfort tradeoff. Readers should inspect clothing, metabolic and environmental assumptions, and keep the chosen comfort index distinct from every occupant's subjective response.
Architectural question
How does a physically grounded reward change a control policy?
Towards an automated AI-based framework for floor plan compliance checks for residential buildings
RESEARCH /
The proposed framework addresses apartment-plan assessment under Australian design policies, including geometric and spatial requirements. It studies automation of a fragmented checking process. Architectural readers should retain the policy version and jurisdiction, and separate detected features from a complete legal or planning-compliance determination.
Architectural question
Which policy-specific assumptions should a plan checker expose?
Subash Gautam, Debaditya Acharya, Alexandra Kleeman, Sarah Foster
HypergraphFormer: Learning Hypergraphs from LLMs for Editable Floor Plan Generation
RESEARCH /
HypergraphFormer uses a language model to generate a textual hypergraph of spatial relationships and connections. It studies editable floor-plan representations across datasets. Architectural review should test whether modifications preserve required relationships and whether the textual structure remains consistent with the resulting geometry.
Architectural question
Can a hypergraph keep layout relationships consistent during editing?
Nikita Klimenko, Hesam Salehipour, Parham Eftekhar, Amir Khasahmadi, Ramon Elias Weber
Generative Floor Plan Design with LLMs via Reinforcement Learning with Verifiable Rewards
RESEARCH /
This paper fine-tunes a language model on plans and uses verifiable rewards to improve numerical and topological constraint adherence. Its architectural focus is dimensions, areas and room connectivity. Readers should inspect the reward definitions and rejected examples, because measurable constraints do not exhaust functional design quality.
Architectural question
What design requirements remain outside a verifiable reward?
Luis Lara, Aristides Milios, Zhi Hao Luo, Aditya Sharma, Ge Ya Luo, Christopher Beckham, Florian Golemo, Christopher Pal
VLM-based automatic multi-granularity graph representation of building layouts for design informatics
RESEARCH /
The authors propose several graph granularities for public-building layouts and use a vision-language model to construct them. It supports retrieval, reasoning and BIM enrichment questions. A practical review should verify node identification, inferred links and information lost when a detailed graph is coarsened.
Architectural question
Which layout relationships survive graph coarsening?
LLM-Guided Agentic Floor Plan Parsing for Accessible Indoor Navigation of Blind and Low-Vision People
RESEARCH /
The proposed framework turns a floor-plan image into a spatial knowledge graph and navigation instructions for blind and low-vision users. Its relevance is accessibility-oriented interpretation. Any real deployment would need user-centered validation and checks for obstacles, changed layouts and hazards that the plan cannot reveal.
Architectural question
What must be verified before plan-based instructions support accessible navigation?
Agentic AI-Enabled Framework for Thermal Comfort and Building Energy Assessment in Tropical Urban Neighborhoods
RESEARCH /
The proposed Singapore-oriented framework links language-model reasoning with lightweight thermal and airflow models. It investigates converting urban questions and policies into evaluation steps. Readers should distinguish a proposed reasoning loop from validated neighborhood predictions and examine the simplified physics used by each activated model.
Architectural question
Which physics checks should accompany an agent's urban-design reasoning?
UrbanFlow-3K: A Dataset of 3,000 Lattice-Boltzmann Simulations of Random Building Layouts
RESEARCH /
UrbanFlow-3K supplies simulated flow fields around randomized building layouts for machine-learning research. It is a training and benchmark resource, not a substitute for project-specific wind assessment. Architectural readers should inspect dimensional assumptions, boundary conditions and whether their own urban geometry falls within the dataset's scope.
Architectural question
Which simulation assumptions limit a learned urban-flow predictor?
Hojin Lee, Andreas Lintermann, Sangseung Lee, Mario Rüttgers
Carbon-aware Market Participation for Building Energy Management Systems
RESEARCH /
The paper combines optimization with transformer forecasts of prices and carbon intensity for building-level energy management. It studies coordinated imports, storage and flexible demand. Architectural and facilities interpretation should expose the assumed market data and examine energy, carbon and comfort consequences separately.
Architectural question
How should a building controller balance price forecasts and carbon signals?
Directly from Alpha to Omega: Controllable End-to-End Vector Floor Plan Generation
RESEARCH /
This study explores direct, controllable generation of residential layouts rather than a fixed chain of intermediate representations. It frames topology, geometry and appearance as connected requirements. Architectural review should test controls against a written brief and inspect failures when the input differs from familiar training examples.
Architectural question
How does an end-to-end plan generator respond to unfamiliar briefs?
3D-RE-GEN: 3D Reconstruction of Indoor Scenes with a Generative Framework
RESEARCH /
3D-RE-GEN reconstructs a scene into textured objects and a background from a single image. It addresses editability and spatial relationships rather than image quality alone. For architectural use, generated object placement and hidden surfaces need explicit review before a reconstructed scene becomes design evidence.
Architectural question
Which reconstructed scene components are genuinely editable and spatially credible?
Tobias Sautter, Jan-Niklas Dihlmann, Hendrik P. A. Lensch
Enhancing Floor Plan Recognition: A Hybrid Mix-Transformer and U-Net Approach for Precise Wall Segmentation
RESEARCH /
MitUNet combines transformer context with a U-Net-style decoder to recognize walls, including thin structures. The study links segmentation precision with reconstruction from drawings. Architectural review should separately examine wall continuity, thickness, openings and symbols so that a plausible mask does not become an inaccurate three-dimensional shell.
Architectural question
Which wall errors matter most when reconstructing a plan in three dimensions?
Floor Plan-Guided Visual Navigation Incorporating Depth and Directional Cues
RESEARCH /
The proposed navigation framework combines a plan with visual information while addressing differences between the two representations. It adds depth and directional cues to the alignment problem. Architectural relevance lies in understanding which plan features remain useful when the observed environment is incomplete or changed.
Architectural question
Which cues help align a plan with an observed interior?
LayoutGKN: Graph Similarity Learning of Floor Plans
RESEARCH /
LayoutGKN studies comparisons between floor plans represented as graphs, aiming to reduce costly interactions during matching. It supports architectural retrieval and clustering questions. Readers should ask whether faster comparison preserves meaningful distinctions between circulation, room relationships and superficially similar layouts.
Architectural question
Can efficient graph comparison retain meaningful layout differences?
Casper van Engelenburg, Jan van Gemert, Seyran Khademi
Extracting Structured Requirements from Unstructured Building Technical Specifications for Building Information Modeling
RESEARCH /
This research studies entity and relationship extraction from French building specifications using language models and comparison methods. It addresses the passage from prose to structured requirements. A practical review must retain source clauses, language context and ambiguous requirements rather than silently resolving them.
Architectural question
Can extracted BIM requirements remain traceable to their source clauses?
Insaf Nahri, Romain Pinquié, Philippe Véron, Nicolas Bus, Mathieu Thorel
Eliminating Rasterization: Direct Vector Floor Plan Generation with DiffPlanner
RESEARCH /
DiffPlanner challenges workflows that repeatedly convert floor plans between vectors and raster images. It studies direct generation within a boundary. Architectural readers can examine whether avoiding rasterization preserves dimensional and topological information, while still requiring checks of room requirements, circulation and editable output.
Architectural question
What information does direct vector generation preserve?
A Scalable Machine Learning Pipeline for Building Footprint Detection in Historical Maps
RESEARCH /
The authors propose a scalable machine-learning pipeline for sparsely distributed buildings in historical rural maps. It supports questions about earlier settlement patterns. Before comparing periods, researchers must account for map conventions, missing sheets and different levels of cartographic detail.
Architectural question
How can historical map conventions distort settlement comparisons?
Physics-Guided Memory Network for Building Energy Modeling
RESEARCH /
This model combines neural learning with predictions from physics-based building simulations. It addresses cases with limited historical data, including newer buildings. Architectural readers should compare the physical-model assumptions with the actual design and check whether their inaccuracies propagate through the learned forecasting system.
Architectural question
Can simulation guidance help a new building without operational history?
Muhammad Umair Danish, Kashif Ali, Kamran Siddiqui, Katarina Grolinger
SCANet: Split Coordinate Attention Network for Building Footprint Extraction
RESEARCH /
SCANet proposes a split-coordinate attention mechanism for footprint extraction. The architectural reading question is how the method handles buildings of different shapes and scales. A clearer segmentation map remains a research output until its boundaries are checked against local survey requirements.
Architectural question
How should footprint extraction be tested across building scales?
Quantifying Key Design Factors for Thermal Comfort in Underground Space Through Global Sensitivity Analysis and Machine Learning
RESEARCH /
Using a Fuzhou underground-space case, the authors combine measured conditions, sensitivity analysis and machine learning to examine thermal-comfort factors. The case is useful for comparing air, radiation and humidity effects. Its reported conditions do not establish a universal underground cooling strategy.
Architectural question
Which comfort variables should an underground-space study measure?
Shisheng Chen, Nyuk Hien Wong, Chao Cen, Ruohan Xu, Lei Xu, Zhenjiang Shen, Zhigang Wu, Jiayan Fu, Zhongqi Yu
BuildEvo: Designing Building Energy Consumption Forecasting Heuristics via LLM-driven Evolution
RESEARCH /
BuildEvo uses language models in an evolutionary process to develop energy-prediction heuristics informed by building characteristics and operations. Its relevance is interpretability alongside forecasting. A review should inspect the generated rule, its physical plausibility and performance on buildings outside the development dataset.
Architectural question
Can a generated forecasting rule remain interpretable and transferable?
