Automated Pipeline Achieves Graph-Based Intelligence for Performance-Driven Building Design

Researchers are tackling the challenge of integrating artificial intelligence with early-stage building designs, which typically lack the structured data needed for effective analysis. Jun Xiao, Qiong Wang, and Yihui Li from the Department of Building Science at Tsinghua University, alongside Zhexuan Yu, Hao Zhou et al, present a fully automated pipeline that converts flexible, unstructured building geometry into intelligent, graph-based Building Information Models and executable Building Energy Models. This innovative framework uniquely bridges the gap between initial design concepts and performance-driven optimisation, allowing AI to understand building elements and their thermal properties with unprecedented clarity. By automating geometry cleansing, space generation, and knowledge modelling, this research significantly expands the potential of BIM and graph-based intelligence to encompass the crucial early phases of architectural design, paving the way for more sustainable and efficient buildings.

The research addresses a critical gap between flexible, early-stage design representations and the structured data required for effective AI reasoning in construction. This innovative approach bridges the divide between initial boundary-representation (B-rep) models, typically lacking explicit spatial or semantic information, and the detailed knowledge graphs underpinning modern AI applications. The team achieved this by integrating automated geometry cleansing, advanced space-generation strategies, and graph-based extraction of spatial and element topology, culminating in a reversible transformation between ontology-based BIM and EnergyPlus energy models.

The study unveils a robust pipeline that begins with unstructured design models and culminates in performance-ready data, enabling AI to interpret building elements, spatial relationships, and associated thermal characteristics. Researchers meticulously combined several key techniques, including automated geometry cleansing to prepare the initial B-rep data, followed by multiple auto space-generation strategies to define interior volumes. Crucially, the framework employs graph-based extraction to define space and element topology, then aligns this with established ontologies to create a knowledge model, effectively translating geometric data into semantic information. Validation using parametric, sketch-based, and real-world building datasets demonstrated the framework’s high robustness and consistent topological reconstruction, ensuring reliable performance-model generation.
Experiments show that this framework not only accurately reconstructs building topology but also reliably generates performance models, demonstrating its potential to significantly enhance early-stage design workflows. This capability unlocks performance-driven design exploration and optimization through learning-based methods, allowing architects and engineers to evaluate design options more effectively and efficiently. The work opens new avenues for AI-assisted design, enabling automated analysis and optimisation of building performance during the crucial early stages of the design process.

This breakthrough reveals a pathway to leverage the power of AI in architectural design by addressing the limitations of current workflows that struggle with unstructured data. Validation across diverse datasets, including parametric models, sketches, and real-world buildings, confirms the framework’s adaptability and reliability, suggesting broad applicability in various architectural contexts. This innovative pipeline addresses the critical representation gap between early-stage design data and the structured formats required for artificial intelligence reasoning. The research team engineered a system that first cleanses the input B-rep geometry, preparing it for subsequent analysis and ensuring data integrity. Following this, multiple automated space-generation strategies were implemented to define and delineate spaces within the building design, crucial for thermal and performance modelling.
Experiments employed graph-based extraction techniques to meticulously reconstruct space and element topology from the cleaned geometry, establishing the relationships between different building components. This topological reconstruction is consistently reliable, demonstrated through validation on parametric, sketch-based, and real-world building datasets. The study pioneered an ontology-knowledge modelling approach, leveraging established ontologies to imbue the BIM with semantic meaning and facilitate AI interpretation. A reversible transformation process was then developed, enabling seamless conversion between the ontology-based BIM and EnergyPlus energy models, a widely used building performance simulation engine.

Scientists harnessed this bidirectional capability to create a closed-loop system where design changes can be rapidly assessed for their impact on building performance. The framework achieves high robustness in handling diverse input geometries, consistently reconstructing topological relationships, and generating reliable performance models. Validation across the three dataset types, parametric, sketch-based, and real-world, demonstrated the system’s adaptability and accuracy. This method enables AI to explicitly interpret building elements, spatial topology, and their associated thermal and performance attributes, unlocking new possibilities for performance-driven design exploration and optimisation.

The approach delivers an AI-oriented infrastructure extending BIM- and graph-based intelligence pipelines to flexible early-stage design geometry, a significant advancement over existing methods limited to well-structured BIM datasets. By bridging design models, BIM, and BEM, the research team created a system that addresses the challenge of applying AI to the earliest phases of architectural design, where models are often unstructured and lack semantic information. The research successfully integrates automated geometry cleansing, space generation, graph-based topology extraction, and ontology-knowledge modeling, culminating in a reversible transformation between ontology-based BIM and EnergyPlus energy models. Validation utilising parametric, sketch-based, and real-world building datasets demonstrated high robustness, consistent topological reconstruction, and reliable performance-model generation. This breakthrough delivers a crucial infrastructure extending BIM and graph-based intelligence pipelines to flexible early-stage design geometry, enabling performance-driven design exploration and optimisation via learning-based methods.

Experiments revealed exceptional energy model accuracy, with lighting and appliance results showing near-zero Root Mean Squared Error (RMSE) and R2 values approaching 1.0. Heating and cooling load predictions achieved monthly R2 values of 0.946 and 0.987, respectively, alongside RMSE values of 0.51 and 0.50 kWh/m2, demonstrating high geometric and thermal fidelity. Measurements confirm that any discrepancies beyond EnergyPlus numerical noise stemmed from ground-connection inconsistencies and minor face standardisation during the BIM-to-BEM stage. The team measured a maximum RMSE of less than 10−12 kWh/m2, attributable to EnergyPlus numerical noise, highlighting the precision of the transformation process.

Data shows that full automatic cleansing increased average robustness from 89% to 100% and spatial accuracy from 71.9% to 81.4%. The cleansing module consistently improved both accuracy and efficiency, reducing problematic geometries encountered during topology generation. Tests prove that even automatic transform, preserving B-rep geometry exactly, achieved a 95% success rate with an average spatial accuracy of 73.4%, despite relying on input quality. Enhanced automatic transform ensured full robustness, while the enhanced semi-automatic workflow achieved an average spatial accuracy of 82% with no failures.

Researchers recorded significant performance gains when embedding the framework within MOOSAS and comparing it to SEFAIRA on the SketchUp platform. In complex high-rise buildings, SEFAIRA experienced delays during cleansing, whereas the proposed framework maintained high stability across all cases, with a maximum spatial accuracy error of 22%. Notably, in a particularly complex case, Fallingwater Apartment, GFA deviation reached 638% for SEFAIRA, indicating a failure to filter redundant elements, while the framework maintained consistent performance. The team measured a maximum spatial accuracy (sA) of 0. This automated process enables artificial intelligence to interpret building elements, spatial arrangements, and associated performance characteristics with greater clarity. The framework integrates geometry cleansing, automated space generation, graph-based topology extraction, and ontology-based knowledge modelling, allowing for reversible conversion between BIM and EnergyPlus energy models. Researchers validated this framework using parametric, sketch-based, and real-world building datasets, demonstrating its robustness and consistent reconstruction of building topology.

The system achieved a 100% processing success rate and an average space-related topology accuracy of 0.871, producing analytical BEMs suitable for performance-informed AI applications. By bridging flexible design geometry with graph-centric AI representations, this work extends the capabilities of knowledge-graph-based BIM techniques to unstructured design models, potentially unlocking applications like performance prediction using graph neural networks and reasoning with knowledge-graph-enhanced large language models. The authors acknowledge limitations regarding computational efficiency and the presence of geometric artifacts, suggesting future work will focus on these areas. Further enhancements to thermal and functional zoning are also planned to strengthen the framework’s utility as an AI infrastructure for early-stage design intelligence. This research offers a significant step towards automating performance-driven design exploration and optimisation through learning-based methods, facilitating reproducible and scalable automation in construction informatics.