AI-Based Methods for Freeform Architectural Geometry: Structural Optimization, Reuse, and Semantic Understanding

This thesis explores how artificial intelligence techniques can support the design, optimization, and understanding of free-form architectural geometries, with a particular focus on gridshells and thin shell structures. While contemporary digital design tools enable the creation of highly expressive free-form surfaces, ensuring their structural feasibility, material efficiency, and sustainability remains challenging, especially when dealing with the preservation of existing designs. At the same time, recent advances in 3D deep learning and multimodal foundation models offer new opportunities for reasoning about geometric data, although their applicability to architectural free-forms is still limited. The thesis first establishes a methodological foundation by reviewing 3D shape representations and learning paradigms in relation to architectural geometry. This analysis highlights the limitations of commonly used geometry processing algorithms and learning backbones when applied to non-watertight surface patches, and motivates the use of mesh-based geometric deep learning as a suitable framework for physics-aware and design-preserving tasks. Building on this foundation, the thesis introduces a learning-based approach for statics-aware shape optimization of free-form gridshells. The proposed method formulates the problem as a single-instance learning task, in which a neural network is trained directly on a single input geometry to iteratively learn vertex displacements that reduce structural strain energy while preserving the original design intent. By integrating differentiable statics-related objectives into an end-to-end neural pipeline, the approach avoids the need for large datasets or explicit ground-truth targets. The framework is then extended to address sustainability-driven constraints through reuse-oriented design. By combining discrete inventory assignment with continuous geometry optimization, composite pipelines are developed to maximize the reuse of reclaimed structural elements while maintaining structural feasibility and geometric coherence. Interactive tools are provided to integrate these methods into established architectural workflows. In addition to performance-driven modification, the thesis investigates semantic understanding of free-form architecture. A brief prompting activity reveals the limitations of current general-purpose foundation models in this field. To overcome the lack, a domain-specific dataset of architectural shells paired with detailed textual descriptions is constructed, and shared latent representations aligning 3D geometry with natural language are learned. These embeddings enable semantic retrieval and exploration of architectural freeforms. Overall, the thesis demonstrates how AI can act as an assistive design technology that operates directly on geometric representations, bridging structural optimization, sustainability, and semantic understanding.