Graph neural networks for fast structural analysis to support interactive design decision-making

This study presents a graph-based surrogate modeling framework for fast prediction of structural responses to support the preliminary design of multi-story buildings under static wind loads. Structures are encoded as graphs, with joints as nodes and beams and columns as edges, enriched by topological and physical attributes. Message-passing Graph Neural Networks (GNNs) are trained on a synthetic dataset to learn a predictive mapping from structural form, load paths, and nodal structural responses. The model predicts displacements, shear forces, and bending moments with high accuracy and demonstrates promising generalization for taller, unseen configurations. Once trained, the GNNs provide near-instantaneous predictions, enabling rapid exploration of design alternatives. To explore its educational potential, this model is also integrated into a Rhino-Grasshopper environment to support early-stage design workflows and interactive feedback. This approach offers a fast and scalable alternative to traditional simulations, helping users make more informed decisions during conceptual design.