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 3 / 5
Tell2Design: A Dataset for Language-Guided Floor Plan Generation
RESEARCH /
Tell2Design pairs floor plans with language instructions to support research on text-conditioned design. The authors emphasize that designs carry spatial constraints beyond those of artistic images. It is useful for examining whether descriptions specify relationships clearly enough to become measurable generation requirements.
Architectural question
Which language instructions can become testable spatial constraints?
Sicong Leng, Yang Zhou, Mohammed Haroon Dupty, Wee Sun Lee, Sam Conrad Joyce, Wei Lu
Data-driven building energy efficiency prediction using physics-informed neural networks
RESEARCH /
This work introduces a physics-informed network using building information, audited characteristics and heating-consumption data. It addresses envelope-related energy prediction. A design review should ask how physical assumptions enter the model and whether audited inputs provide enough evidence to distinguish envelope losses from operational effects.
Architectural question
Can a learned energy model separate envelope effects from operation?
Generative AIBIM: An automatic and intelligent structural design pipeline integrating BIM and generative AI
RESEARCH /
Generative AIBIM proposes a structural-design pipeline connecting BIM and generative methods, including physical conditioning. It addresses how design generation can respond to engineering context. Architectural adoption would still need independent load, structural and coordination checks before generated arrangements become a project design.
Architectural question
Which engineering conditions must constrain a generated structural scheme?
TFNet: Tuning Fork Network with Neighborhood Pixel Aggregation for Improved Building Footprint Extraction
RESEARCH /
TFNet addresses closely connected buildings and weak predictions at image boundaries by incorporating neighborhood context. It is relevant to continuous urban mapping across image tiles. The practical review should examine whether stitching produces missing edges, merged buildings or inconsistent polygons across adjacent captures.
Architectural question
How should footprint maps remain consistent across image tiles?
Muhammad Ahmad Waseem, Muhammad Tahir, Zubair Khalid, Momin Uppal
Terrain-Informed Self-Supervised Learning: Enhancing Building Footprint Extraction from LiDAR Data with Limited Annotations
RESEARCH /
This paper uses terrain information to support self-supervised building-footprint learning with limited labels. It addresses annotation scarcity and transfer across remote-sensing settings. Architectural mapping should inspect how slopes, ground classification and local terrain assumptions influence the boundary ultimately assigned to each building.
Architectural question
How does terrain information change LiDAR-based footprint inference?
Anuja Vats, David Völgyes, Martijn Vermeer, Marius Pedersen, Kiran Raja, Daniele S. M. Fantin, Jacob Alexander Hay
MARL: Multi-scale Archetype Representation Learning for Urban Building Energy Modeling
RESEARCH /
MARL uses learned geometric representations to construct building archetypes for urban energy modeling. It questions whether nationally defined types overlook local form. Its architectural relevance is choosing representative models that retain meaningful geometry without pretending that every individual building has been measured in detail.
Architectural question
How much local geometry should an urban energy archetype preserve?
Skip-Connected Neural Networks with Layout Graphs for Floor Plan Auto-Generation
RESEARCH /
The proposed network combines layout graphs with multiscale image information to generate plans. It reports a benchmark challenge context and provides a code link. Architectural assessment should look beyond segmentation scores to inspect access, room proportions and whether generated arrangements can be edited coherently.
Architectural question
What should be checked beyond a generated plan's pixel score?
Enhancing personalised thermal comfort models with Active Learning for improved HVAC controls
RESEARCH /
This study uses active learning to select informative moments for occupant feedback instead of collecting preferences indiscriminately. It addresses the burden of personal comfort data collection. A useful pilot should compare annotation effort, missed discomfort and the stability of the resulting control model.
Architectural question
Which feedback moments are worth asking occupants to annotate?
Zeynep Duygu Tekler, Yue Lei, Xilei Dai, Adrian Chong
SSIG: A Visually-Guided Graph Edit Distance for Floor Plan Similarity
RESEARCH /
SSIG proposes a graph-edit-based measure of floor-plan similarity without training a model. It is relevant to evaluating AI-generated layouts and retrieving precedents. The architectural question is whether its distance reflects room shapes and relationships that matter to a designer, rather than only image resemblance.
Architectural question
Which similarity metrics fairly evaluate generated plans?
Casper van Engelenburg, Seyran Khademi, Jan van Gemert
Building Footprint Extraction in Dense Areas using Super Resolution and Frame Field Learning
RESEARCH /
The proposed pipeline enhances aerial-image resolution and combines segmentation with frame-field learning for dense areas. Its architectural relevance is separating irregular, closely spaced buildings. Upsampled detail must be treated cautiously: a sharpened image can improve interpretation without creating new measured information.
Architectural question
Where can super resolution mislead dense-building interpretation?
Expediting Building Footprint Extraction from High-resolution Remote Sensing Images via progressive lenient supervision
RESEARCH /
The authors investigate a training strategy intended to ease the transfer of feature extractors into building-footprint segmentation. It addresses computation and difficult boundary regions. Urban-context users should examine what is gained in transfer efficiency and whether ambiguous edges remain accurately represented in the final map.
Architectural question
Can a lighter training strategy retain difficult building boundaries?
Haonan Guo, Bo Du, Chen Wu, Xin Su, Liangpei Zhang
Building Footprint Extraction with Graph Convolutional Network
RESEARCH /
The authors use graph-based learning to address fine-grained boundaries that can be lost in conventional convolutional segmentation. It contributes to vector-quality urban mapping questions. Architectural users should inspect edge fidelity and building separation in representative local imagery before using predicted footprints as project context.
Architectural question
Can graph learning preserve the boundaries needed for urban-context drawings?
Using Text-to-Image Generation for Architectural Design Ideation
RESEARCH /
The authors report a laboratory study in which architecture students developed a cultural-center concept with image generators. It investigates creativity and constraint handling rather than completed project performance. The case is useful for discussing where generated references help ideation and where they distract from a specific brief.
Architectural question
When do generated images help a concept rather than distract from its brief?
BCE-Net: Reliable Building Footprints Change Extraction based on Historical Map and Up-to-Date Images using Contrastive Learning
RESEARCH /
BCE-Net compares historical footprints with newer images using contrastive learning. It addresses false change signals from seasonal differences and viewing geometry. Urban monitoring should separate confirmed building change from changed appearance, and retain uncertain cases for source-level inspection before updating a database.
Architectural question
How can a map update distinguish demolition from changed imagery?
SuperpixelGraph: Semi-automatic generation of building footprint through semantic-sensitive superpixel and neural graph networks
RESEARCH /
This approach combines semantically sensitive superpixels and graph learning for footprint extraction with human involvement. It addresses the effort of correcting overly smooth polygons. A useful architectural mapping pilot should record edit time and retained geometry, rather than counting automatic polygons as completed survey information.
Architectural question
When does semi-automatic extraction reduce real correction work?
Haojia Yu, Han Hu, Bo Xu, Qisen Shang, Zhendong Wang, Qing Zhu
Attention Boosted Autoencoder for Building Energy Anomaly Detection
RESEARCH /
The study models normal consumption behavior with an attention-assisted autoencoder and looks for deviations. It is relevant to identifying operational problems early. A building team should distinguish meaningful faults from schedule changes, maintenance and legitimate occupancy variation before an anomaly becomes a corrective action.
Architectural question
Which energy anomalies represent faults rather than legitimate changes?
Parsing Line Segments of Floor Plan Images Using Graph Neural Networks
RESEARCH /
GLSP predicts endpoints and classifies line segments using a graph network, producing vectorized elements from plan images. It offers an alternative to pixel-only recognition. Architectural document conversion should inspect connections, gaps and line meanings before the extracted vectors are reused as walls or openings.
Architectural question
What checks make extracted plan vectors usable for modelling?
The method introduces pseudo-plane regularization for neural surface reconstruction in low-texture indoor regions. It focuses on walls and floors that image-based methods can struggle to recover. Architectural users should inspect whether these geometric priors help useful surfaces without erasing intentional curvature or irregularity.
Architectural question
When can planar priors distort a room's actual geometry?
BiSVP: Building Footprint Extraction via Bidirectional Serialized Vertex Prediction
RESEARCH /
BiSVP represents each building as an ordered vertex sequence and predicts it in two directions. It offers a vector-oriented alternative to mask extraction followed by refinement. The design-use question is how corners, ordering and small separations survive direct geometric prediction.
Architectural question
Can direct vertex prediction produce a trustworthy contextual polygon?
Mingming Zhang, Ye Du, Zhenghui Hu, Qingjie Liu, Yunhong Wang
AutoTherm: A Dataset and Benchmark for Thermal Comfort Estimation Indoors and in Vehicles
RESEARCH /
AutoTherm compares thermal-comfort estimation in buildings and vehicles and provides a temporal benchmark. It highlights transfer limits between relatively stable rooms and changing conditions. Architectural use is the indoor component; vehicle results should not be imported as proof of building comfort-model reliability.
Architectural question
What changes when comfort models move from static rooms to dynamic conditions?
Mark Colley, Sebastian Hartwig, Albin Zeqiri, Timo Ropinski, Enrico Rukzio
Meta-Reinforcement Learning for Building Energy Management System
RESEARCH /
The paper explores meta-learning to address the training burden of reinforcement-learning energy management. It asks how experience can transfer between control situations. Architectural adoption should examine the limits of that transfer, including changed equipment, occupancy and weather, rather than assuming previously learned behavior remains suitable.
Architectural question
When can energy-control experience transfer between buildings?
Huiliang Zhang, Di Wu, Arnaud Zinflou, Benoit Boulet
NeuralRoom: Geometry-Constrained Neural Implicit Surfaces for Indoor Scene Reconstruction
RESEARCH /
NeuralRoom reconstructs room-scale surfaces from images while treating textured and textureless areas differently. The paper addresses ambiguity between shape and appearance. For architectural work, its main reading question is whether flat walls and complex details remain geometrically credible in the same reconstruction.
Architectural question
How should a reconstruction handle both plain walls and detailed objects?
Are You Comfortable Now: Deep Learning the Temporal Variation in Thermal Comfort in Winters
RESEARCH /
This study asks how time of day, circadian rhythm and outdoor temperature affect learned thermal-comfort predictions. It challenges models that treat responses as static. Architectural application should consider repeated measurements and different operating periods rather than assuming one snapshot represents a person's continuing preference.
Architectural question
How should comfort prediction account for time-dependent preferences?
Building Matters: Spatial Variability in Machine Learning Based Thermal Comfort Prediction in Winters
RESEARCH /
The study examines spatial variation in learned comfort predictions, particularly in naturally ventilated buildings. It questions transferring models designed for mechanically conditioned adult environments. Architectural interpretation should consider room location, local conditions and the population being studied rather than assuming one building-wide prediction is adequate.
Architectural question
When is one building-wide comfort model too coarse?
