LLM4Laser: Large Language Models Automate the Design of Lasers

TL;DR

The paper investigates using large language models to automate PCSEL design through a human-AI co-design workflow, enabling concept brainstorming, code generation, and RL-driven optimization of photonic structures. By enabling GPT-4 to draft FDTD (Meep) simulations and DQN-based optimization scripts, the approach aims to create an end-to-end, low-human-in-the-loop design pipeline that can produce high-quality PCSELs with targeted metrics. The key contributions include a practical three-stage workflow, five practical prompts (golden tricks) for effective LLM collaboration, and demonstration of a DQN-based design optimization achieving a wavelength near 1310 nm with $Q\ge 10^4$, a lasing area $\ge 0.36~\mu m^2$, and divergence $\le 3^\circ$, along with a comparative assessment against open-source LLMs. The work demonstrates a significant step toward self-driving laboratories in nanophotonics, while outlining limitations (e.g., infinite-area simulations) and future directions for broader metrics and tape-out readiness.

Abstract

With the rapid evolution of global autonomous driving technology, the demand for its core sensing hardware, Light Detection and Ranging (LiDAR), is escalating. As the light source part of the LiDAR system, lasers, particularly the cutting-edge Photonic Crystal Surface Emitting Lasers (PCSEL), have correspondingly attracted extensive research attention. The conventional manual design and optimization of PCSEL typically require expertise in semiconductor physics and months of dedicated effort to achieve satisfactory results. While AI-driven approaches can expedite this process, laser designers still need to invest time in learning the AI algorithms involved. Meanwhile Large Language Models (LLMs), leveraging their powerful reasoning abilities, can effectively comprehend natural language and provide constructive feedback in multi-turn dialogues. They have already demonstrated potential to assist humans in scientific fields such as robotics design and chemical discovery. A question naturally arises is: Can LLMs transform the lasers design process? This paper proposes a novel human-AI co-design paradigm to show that LLMs can guide the laser design and optimization process both conceptually and technically. Specifically, by simply having conversations, GPT assisted us with writing both Finite Difference Time Domain (FDTD) simulation code and deep reinforcement learning (RL) code to acquire the optimized PCSEL solution, spanning from the proposition of ideas to the realization of algorithms. Given that GPT will perform better when given detailed and specific prompts, we break down the PCSEL design problem into a series of sub-problems and converse with GPT by posing open-ended heuristic questions rather than definitive commands. We achieved a significant milestone towards self-driving laboratories, that is, a fully automated AI-driven pipeline, for laser design and production.

Figures (10)

• Figure 1: Long-term vision of this work: LLMs for automated PCSEL design and optimization pipeline that enables self-driving laboratories. The human facilitator prompts the LLM to generate FDTD code for simulating the PCSEL structure and AI (e.g., reinforcement learning (RL)) code for subsequent optimizations of the PCSEL model. The FDTD code is written with the MIT meep oskooi2010meep package. The AI optimization process with RL is built upon an earlier work's L2DO framework li2023deep. The final optimized PCSEL design (shown on far right) is then converted to CAD layout and prepared for tape-out and foundry fabrication.
• Figure 2: Photonic Crystal Surface Emitting Laser (PCSEL), with abundant applications in sensing, LiDAR, and telecommunications.
• Figure 3: LLM4Laser: A novel Human-AI co-design paradigm for applying LLMs to PCSEL design and optimization. A pictorial overview of the discussions and interactions between the human facilitator and the LLM, with the questions prompted by the human and the answers/solutions provided by the LLM (GPT). The process is divided into three steps: left column: conceptualization, middle column: code generation and debugging, and right column: simulation and optimization. Optimization via DQN is run on high-performance computing (HPC) clusters for improved computational speed and output.
• Figure 4: Code generated by GPT-4 for FDTD simulation of PCSEL using the meep package. Left: geometry setup section, right: simulation setup and calculations section. The code shown here is the final version that runs successfully after several rounds of debugging.
• Figure 5: Learning curves of training the DQN to optimize PCSEL, plotted as scores vs. episodes. (a) Average score of each episode; (b) Maximum score of each episode. Each episode contains a horizon of 500 steps. Vertical axes are plotted in log scale.