The urgent need to address the environmental impacts of the built environment has led to a significant increase in the demand for sustainable buildings. Integrating sustainable performance considerations early in the design process is essential, as changes made at this stage incur considerably lower costs over the entire life cycle of buildings. Early design is inherently complex and involves multiple interrelated factors. Performance-based Generative Design (PGD) is a robust generative method that alleviates the complexities of early design by employing computational techniques to generate design cases across various parameters and shapes while incorporating sustainable performance considerations. However, the current PGD framework imposes substantial computational demands, and excessive time consumption during the early design stages is unacceptable. This limitation hinders the widespread adoption of PGD among architects in real-world projects. Consequently, there is an urgent need to streamline the current PGD framework to reduce computational demands to an acceptable level. Four key areas have been identified for enhancing the PGD framework: 1) the development of an efficient data model for integrating data from multiple information sources in PGD and facilitating automatic simulation modeling; 2) the establishment of a knowledge-informed optimization framework that can effectively narrow solution spaces and reduce computational time; 3) the provision of real-time indicator predictions in sustainable performance evaluation, which can offer prompt feedback to architects during the early design stage; and 4) the development of a general generative framework to achieve agile and creative generation at neighborhood levels. To address these critical issues, this research leverages knowledge graph (KG) technologies to enhance the accuracy and speed of PGD. Firstly, a KG-based automatic framework is proposed to integrate multiple data sources in PGD and automatically generate building energy simulation models. A KG schema for PGD (PGD-KG schema) is developed to encompass data from various domains, including weather, building specifications, internal heat gains, HVAC systems, and regulatory requirements, thereby representing the topological relationships within PGD. Additionally, a cross-domain and rule-based reasoning method for thermal zoning is developed, along with an automatic translation method from PGD-KG models to BEM models, utilizing instance-based mapping and dynamic data conversion. Validation of this approach is achieved through the development of a BEM model for one floor of a campus building, demonstrating that the proposed framework meets the modeling precision achieved by manual methods and has the potential to reduce modeling time by over 99% compared to manual modeling. Secondly, a knowledge-informed PGD optimization framework for sustainable buildings is proposed to mitigate time-related issues by integrating KG into PGD. A generation method is introduced for automatically developing PGD-KG models from parametric design models enhanced with semantic information. Furthermore, cross-domain reasoning algorithms are developed to enable automated compliance checking and the selection of cases with superior sustainable performance based on regulatory requirements and sustainable design strategies. The proposed framework is applied to a design project focused on optimizing module layout to minimize cooling energy and maximize daylighting. Results indicate that the proposed framework can generate a satisfactory number of Pareto-optimal solutions while reducing computational time by 73.25% compared to general optimization frameworks. Thirdly, optimal surrogate models that incorporate inter-building effects are developed to predict multiple indicators of sustainable performance for residential blocks at the regional level, leveraging graph neural networks (GNNs). Components at the neighborhood level are included in the PGD-KG schema to represent the geometric features and relationships among buildings in residential block design. A regional dataset is created for model training and testing, utilizing real residential zones in Hong Kong. The surrogate models are developed and evaluated using two types of architectures (individual architectures for specific indicators and an integrative architecture) and three different neural networks (Graph Attention Network, Graph Convolutional Network, and Artificial Neural Network). Results show that the surrogate models employing individual architectures and GAT outperform those using other architectures and neural networks, achieving high prediction accuracy with CV(RMSE) values of 11.79%, 7.63%, and 8.00% for energy consumption, indoor thermal comfort, and daylighting, respectively, on the regional test set. Moreover, these models enable significant acceleration of performance evaluation during the early design stage, achieving a speedup of 243,297 times compared to physics-based simulations. Lastly but not least, a sustainability-oriented generative framework for residential site layout design, characterized by agile and creative features, is proposed. The problem of development potential is formulated, and corresponding solutions are proposed to obtain block combinations that maximize residential units while adhering to planning regulation constraints. A generation method is developed that produces a variety of candidate design cases from block combinations through a cascading process involving algorithms that efficiently narrow the search space. Additionally, a real-time evaluation method is proposed based on knowledge graph-embedding surrogate modeling to identify Pareto-optimal design cases aligned with sustainable goals. The proposed framework is implemented in a site layout design project, generating 4,660 candidate design cases and 49 Pareto-optimal design cases within 2.6 hours, showcasing its agile and creative capabilities. Compared to a benchmark method from recent studies, the proposed generation method increases the number of generated cases by 618% while enhancing design variety and reducing generation time by 81.86%. When compared to physics-based simulations, the proposed evaluation method accelerates the evaluation process by 48,674 times while maintaining acceptable prediction accuracy, with CV(RMSE) values below the maximum acceptance threshold of 25% set by ASHRAE. In summary, this research enhances the accuracy and speed of PGD by leveraging KG technologies. By employing the techniques developed in this study, architects can efficiently navigate a variety of design options toward sustainability within an acceptable timeframe during the early design stages. This facilitates rapid conceptualization, modeling, and feedback, thereby promoting the widespread adoption of PGD among architects in real-world projects and contributing to the realization of sustainable concepts within the Architecture, Engineering, and Construction (AEC) sector.
| Date of Award | 2025 |
|---|
| Original language | English |
|---|
| Awarding Institution | - The Hong Kong University of Science and Technology
|
|---|
| Supervisor | Jack Chin Pang CHENG (Supervisor) & Zhe WANG (Supervisor) |
|---|