Key considerations for robust near-field response prediction and optical metasurface inverse design

Optical metasurfaces, used in applications such as signature management and wavefront sculpting, are increasingly leveraging artificial intelligence (AI) for inverse design. The objective is to significantly accelerate the discovery of innovative solutions, moving beyond traditional, slow, and inefficient simulation-based methods or reliance on human intuition. In this article, various AI-driven approaches are examined for learning a robust forward prediction emulator, a key step in inverse design. Despite their potential, challenges remain, including issues with accuracy, generalization, and the need for unrealistic amounts of data. The networks are compared and contrasted, ranging from direct models that map design parameters to waves to sequence-based approaches that aim to learn incremental wave propagation. Using a dataset generated in MEEP, best practices are sought regarding network architecture, optimization, penalization, evaluation criteria, and experimental design. Ultimately, these details and insights are intended to enhance research reproducibility, address the practical implementation of these networks, and offer recommendations for future AI solutions.