AI-driven metasurface design pipeline combining tandem neural networks with deep reinforcement learning to automate broadband RCS reduction. An LSTM-based policy network generates 4 base unit topologies (7×7 binary), which are expanded to 8 units via symmetry and rotation, then evaluated by a surrogate forward model predicting S11/S22 phase responses across 8–18 GHz. The agent optimizes a 6×6 metasurface arrangement using GRPO, maximizing RCSR (>10 dB) and bandwidth through end-to-end differentiable training with gradient feedback from loss functions combining phase error and RCS reduction objectives. Fork this diagram to customize frequency bands, unit cell dimensions, reward weights, or LSTM hidden states for your own electromagnetic design automation. The tandem architecture—inverse policy model paired with forward surrogate—eliminates expensive EM simulations during training, enabling rapid exploration of topology and arrangement spaces.