IASS Annual Symposium 2025, Mexico City

As architecture faces rising demands for material efficiency, adaptability, and intelligent systems integration, new computational frameworks are needed to align performance with generative design. This paper presents a machine learning framework for predicting structurally meaningful lattice geometries in freeform architectural shells based on stress input. We introduce a Transformer-based model trained on scalar and directional stress fields to infer reinforcement patterns, producing polyhedral lattices aligned with principal stress trajectories. Operating without templates or rule-based encoding, the model generalizes across varied topologies and boundary conditions. By learning neighborhood relationships and stress flows, it internalizes structural logic beyond local cues, generating fabrication-ready outputs. Inference time is reduced from minutes to milliseconds, enabling a new scale of real-time structural reasoning. This approach bridges simulation and design, positioning AI as a tool for adaptive, performance-driven fabrication in architectural practice.