Learning Electromagnetic Metamaterial Physics With ChatGPT

The record describes using large language models (LLMs) to learn electromagnetic metamaterial behavior. Specifically, it focuses on all-dielectric metamaterials made from unit cells with four elliptical resonators. The authors present a fine-tuned LLM trained on up to 40,000 data points that predicts the absorptivity spectrum when given a text prompt describing only the metasurface geometry. Its performance is compared against conventional machine learning methods such as feed-forward neural networks, random forest, linear regression, and K-nearest neighbor. The work also explores inverse problems: using the LLM to suggest the geometry needed to achieve a desired absorptivity spectrum. The abstract argues that LLMs may help research by processing large datasets, finding hidden patterns, and operating in higher-dimensional spaces.