LLM4Laser: Large Language Models Automate the Inverse Design of Lasers

Renjie Li

Abstract: AI-driven inverse design has transformed nanophotonic de- vice discovery, yet remains fundamentally limited: while AI algorithms accelerate optimization, the simulation and training infrastructure must still be hand-crafted by domain experts. Here we demonstrate that large language models (LLMs) can autonomously construct complete inverse- design workflows through natural-language interaction. We guided an LLM to generate electromagnetic simulations, reinforcement-learning optimization, and integrated environ- ments for photonic-crystal surface-emitting lasers (PCSELs). Through iterative dialogue, the LLM refined and debugged these components with minimal human intervention, pro- ducing a functional end-to-end workflow. The resulting au- tonomous framework discovered PCSEL geometries exhibit- ing a Q-factor of 37,000 and normalized radiation loss 18- fold lower than current state-of-the-art devices while main- taining a significantly reduced footprint. To accelerate de- sign further, we fine-tuned a multi-modal LLM to predict op- tical properties directly from device images, enabling rapid performance estimation. Our results establish that LLMs can both construct sophisticated computational infrastructure for nanophotonic inverse design and interpret physical structures visually, providing a pathway toward fully autonomous pho- tonic engineering.