Can deep learning be used effectively in photonic device development? This perspective critically examines the growing emphasis on deep learning frameworks by highlighting persistent challenges in accuracy, data availability, and computational efficiency. Despite their appeal, data-driven, deep learning methods have well-known trade-offs and requirements when compared to numerical techniques, which often make them suboptimal. Furthermore, deep learning methods often succeed with diverse, large datasets that demand significant computational resources for model training. We argue that while deep learning methods may not serve as an immediate replacement in the short term, they may remain valuable for problems where these requirements are already met, particularly as a surrogate to complex design problems and addressing the ill-posed nature of inverse design. Using case studies such as physics-informed neural networks and neural operators, we advocate for an outlook that is optimistic about deep learning’s potential, but pragmatic about when and where it offers genuine advantage over classical techniques.
Iyer et al. (Wed,) studied this question.