Physics-Informed Deep Reinforcement Learning for Automated Section Design with Optimized Seismic Performance in Steel Moment Frame

This study introduces a novel Physic-Informed Deep Reinforcement Learning (PIDRL) for the design in steel moment-resisting frames (SMRF). The selection process incorporates structural geometric properties, code-based design constraints, and multiple earthquake intensity levels, such as Strength Level Earthquake (SLE), Design Basis Earthquake (DBE) and Maximum Considered Earthquake (MCE), in compliance with Performance-Based Design (PBD) requirements. The key innovation lies in the synergistic application of PIDRL, along with machine learning (ML) model to deliver both high accuracy and computational efficiency. After compiling datasets of structural configurations and earthquake records, ML algorithm is first employed to rapidly and accurately predict engineering demand parameters (EDPs) across diverse seismic scenarios, overcoming the time-consuming limitations of traditional numerical simulations. Subsequently, the PIDRL agent optimizes section selection while strictly satisfying functional and regulatory constraints, thus ensuring both performance objectives and minimization of initial construction costs. The outcome is an intelligent section selection tool capable of proposing optimal column sizes for any given steel frame structure. Compared to conventional approaches, this method substantially reduces design time and costs, providing a practical and efficient solution for modern earthquake-resistant structural design.

کلیدواژه ها:

Physics-Informed Deep Reinforcement Learning ، Proximal Policy Optimization ، XGBoost ، Seismic Performance Optimization ، Engineering Demand Parameters