Performance-Driven Design Exploration of Biocomposite Facades

Abstract

With decision-making becoming increasingly data-driven in early design stages due to global environmental challenges, and qualitative metrics remaining crucial in facade design due to their huge architectural impact, performance-driven design exploration frameworks are emerging as powerful tools to explore vast design spaces based on geometry typology and approximated performance. However, those used in facade design based solely on Self-Organising Maps (SOM), face significant challenges in computational efficiency. While more advanced frameworks used in the AEC sector combining SOM and Multi-Layer Perceptrons (MLP) address this, they still face limitations in prediction accuracy, convergence speed, interpretability, reliability, and usability, reducing their effectiveness in decision-making. This thesis aimed to overcome these limitations, by developing a novel framework integrating SOM and Kolmogorov-Arnold Networks (KAN), applied in the design process of an aluminium-based biocomposite curtain wall facade. The results demonstrate that substituting heavy performance simulations with fast approximations using KAN leads to significant enhancements in computational efficiency. In addition, comparative analysis revealed that KAN outperforms MLP in prediction accuracy on highly-complex performance metrics, with much faster convergence and smaller architectures. Furthermore, KAN proved to be faster and more intuitive to train, as well as more consistent in predictions. Moreover, KAN enhanced interpretability between geometry and performance, enabling designers to focus on relevant design variables and adjust them strategically toward optimal performance, providing an integrated solution with transparent decision-making and faster processing compared to traditional sensitivity analysis tools. Finally, a novel approach to design exploration has enabled the integration of less-geometry related design variables, enhancing optimisation capabilities, as well as proven to balance human-AI interaction more efficiently than traditional frameworks, making design exploration more interactive, thereby more effective and intuitive. Ultimately, the SOM-KAN framework has proven to advance the facade design process by facilitating more efficient decision-making in early design stages, leading to superior architectural and sustainable solutions.