AI-guided generation of reticular structures by integrating reinforcement learning with shape grammars

During the early stages of design, there is a need for computational tools that impartially generate and explore diverse structural possibilities. While conventional methods based on the Finite Element Method (FEM) have limited generative capacity, a grammar-based approach—integrating structural information to generate diverse structures—significantly broadens the design space. Navigating this broadened design space involves either sorting and clustering synthetically generated datasets or guiding the system to produce feasible outcomes. This paper introduces a guiding methodology for a grammar-based system, Grammatical Design with Graphic Statics (GDGS), aimed at creating statically equilibrated, goal-oriented 2-dimensional (2D) structures. By incorporating an artificial intelligence (AI)-based technique called reinforcement learning (RL), specifically Q-learning, within this framework, the paper focuses on guiding GDGS to produce diverse yet feasible reticular structures. The synergy between grammar-based rule transformations and RL's action selection forms the foundation for deploying RLbased algorithms with grammatical design methods. The presented experiments demonstrate the potential of incorporating multiple feasibility criteria within the RL system, balancing multiple objectives and exploring alternative latent structural solutions. This approach provides a valuable tool for designers, streamlining the generation of feasible structures. The paper concludes by discussing how RL methods can create synthetic datasets of goal-oriented, high-performance structures, with the sequence of rules acting like encoded DNA, along with performance metrics.