Development of a reinforcement learning-based planner for spatial layout optimization of PEM electrolysis systems

This study proposes a Unity-based multi-agent reinforcement learning (MARL) framework for three-dimensional geometric layout optimization of proton exchange membrane (PEM) water electrolysis systems. Seven major components are optimized using Proximal Policy Optimization (PPO) under collision avoidance, positional, and process-connectivity constraints, with A*-based pipeline routing. The framework achieved average rewards of 0.060–0.080 and pipeline lengths of 0.92–3.89 m while adaptively rearranging equipment under fixed-position constraints. Its scientific novelty lies in integrating PEM-specific 3D geometric constraints, MARL, and automated pipeline routing, while its practical contribution is a simulation-based tool for adaptable layout design extendable to larger systems and constraints.