Towards Large-Scale AI-driven Inverse Design in Metaphotonics: Experimental Generation of Million-Scale Datasets

Metasurfaces provide a versatile platform for the development of modern photonic devices, where inverse design plays a critical role in addressing the complex interplay between their optical responses and geometries [1], [2]. AI-based inverse design models enable fast, accurate predictions of spectrum-geometry relationships, while accounting for fabrication constraints and efficiency requirements [3], [4], surpassing traditional optimization-based methods [5]. However, achieving accurate and well-generalized AI models requires datasets containing millions of training samples - a scale that remains difficult to attain with current electrodynamics simulation methods [3], [5].