UC Berkeley

Structural member sizing directly influences material consumption, embodied carbon, and overall structural efficiency. Recent advances in graph representations have opened new possibilities for structural optimization, with applications such as topology optimization and surrogate modeling. We propose a graph-based representation learning and clustering pipeline that uses structural simulation results to support member sizing and cross-section optimization. In this representation, connection points and linear members in the structural system are modeled as two node types in a heterogeneous graph. Mechanical responses, geometric properties, and topological information are encoded as features, and structural context is propagated through message passing. We then train a Heterogeneous Graph Attention Network (HeteroGAT) encoder with a contrastive objective constructed from mechanically similar and topologically adjacent member pairs. Finally, we cluster the learned embeddings with Gaussian Mixture Models (GMM). The resulting clusters provide a practical basis for structural optimization. Specifically, we use the clustering results as grouping rules for cross-section optimization in Karamba3D, such that members within the same cluster share a common profile. Using steel member sizing as a case study, we evaluate the cross-section assignment strategies on a whole-building structural system and demonstrate the method through multiple application examples. The results show that the proposed method achieves a rationalized section system with significantly fewer cross-section types and stable convergence, while maintaining structural performance comparable to the original design, making it a practical tool for early-stage steel structural design.