From proof of concept to practice: the next decade of artificial intelligence in metasurface photonics

Abstract

Over the past decade, artificial intelligence has learned to predict electromagnetic responses rapidly and generate high-performance metasurface designs. However, one question remains open: compared with conventional design workflows that contain no learned model, has it made the complete design workflow faster, and are the resulting devices better? We argue that the answer depends on the task. Reading 100 highly cited optical studies in depth, we distinguish the design tasks in which artificial intelligence is irreplaceable, clearly advantageous, or without clear advantage. In our view, the key to the next decade is not more successes in isolated tasks but making each success a foundation for the field, through shifts such as shared data, trained models, and physics written into learning. Whether artificial intelligence leaves lasting value in this field will depend on whether its benefits can be verified, its methods transferred, and its designs corrected by experiment.