A physics‐informed deep reinforcement learning framework for autonomous steel frame structure design
Bochao Fu, Yuqing Gao, Wei Wang
Abstract
As artificial intelligence technology advances, automated structural design has emerged as a new research focus in recent years. This paper combines finite element method (FEM) and deep reinforcement learning (DRL) to establish a physics-informed framework, named FrameRL, for automated steel frame structure design. FrameRL models the design process of steel frames as a reinforcement learning (RL) process, enabling the agent to simulate a structural engineer's role, interacting with the environment to learn the methods and policies for structural design. Through computer experiments, it is demonstrated that FrameRL can design a safe and economical structure within 1 s, significantly faster than manual design processes. Furthermore, the design performance of FrameRL is compared with traditional optimization algorithms in three typical design cases and a high-rise steel frame case, demonstrating that FrameRL can efficiently complete structural design based on learned design experiences and policies.