UC Berkeley

Structural design is deeply expertise-driven: engineers rely on intuition, experience, and both explicit and tacit knowledge of load paths, structural typologies, first principles and in-depth calculations to arrive at a design solution. Current computational workflows in structural design largely focus on optimization, form-finding and dimensioning of members, where the design space is predefined and algorithms iteratively refine a solution. Machine learning, however, opens the possibility to explore broader design spaces in which geometry and topology are not fixed in advance. Imitation learning (IL), a paradigm closely related to reinforcement learning, offers a pathway to integrate skills from design examples, rather than explicit rulesets. Recently, flow matching, a generative framework that learns a vector field transporting noise toward data over time, has emerged as a prominent method for modeling complex, multimodal action distributions that standard one-shot predictors often fail to capture. By learning from engineers’ demonstrations with a flow-based imitation policy, we transfer structural intuition into a design agent without dense reward engineering or computationally expensive finite element analysis at every decision step. To study this idea, we create a 2D structural testing environment. Coupled with a finite element analysis solver to track and rank design outcomes, we create a training ground for the IL algorithm. Using the environment we tested building of pin-jointed steel truss bridges as graphs and trained an imitation policy to predict chunks of continuous placement actions and capture longer-term design intent. Our results show that the flow-based policies can learn structural intuition, generating diverse feasible bridge designs that reflect established engineering principles. The work introduces a reproducible benchmark for assembly-constrained structural design and examines the limitation and the potential of generative imitation learning as an alternative framework for structural design exploration beyond top-down optimization.