When Every Simulation Counts: Value-Based Reinforcement Learning for Accelerated Photonics Inverse Design

The paper benchmarks baseline DQN and six value-based variants for a seven-variable photonic-crystal surface-emitting laser (PCSEL) design problem under a shared objective, simulator, 83-call budget, and four matched initializations. Dueling DQN is the only method that improves across all four seeds. Its selected structures reportedly reduce wavelength error by 64% and increase upward power by 47% relative to the first evaluated designs, while achieving a higher mean quality factor than baseline DQN under the same budget. The abstract omits several absolute numerical values, so the magnitude of some claims requires checking the full paper.