AI-driven metasurface design using reinforcement learning with a surrogate model accelerates electromagnetic unit cell optimization. A policy network generates binary configurations for four 7×7 unit cells, which expand via symmetry (S11/S22 swap) into an 8-unit tiled 6×6 array, then fed to a black-box surrogate predictor that estimates S11 and S22 phase responses. RCSR reward signals from phase predictions drive PPO-based policy updates, closing the RL loop without expensive full-wave simulations. This architecture demonstrates how surrogate models reduce computational cost in physics-based design automation. Fork and customize this diagram on Diagrams.so to adapt the symmetry expansion, reward function, or policy network architecture for your metasurface or antenna design workflow.