Research progress on deep learning-based design and application of intelligent micro-nano photonic devices

As fundamental components in optical communications, quantum information processing, and biosensing, micro-nano photonic devices have long faced limitations due to the inherent blindness of traditional trial-and-error approaches and low efficiency of electromagnetic simulation techniques in high-dimensional parameter spaces. This is particularly evident in devices such as metasurfaces, photonic crystal nanocavities and quantum coupling structures, where conventional methods struggle to balance optimisation accuracy with computational efficiency. Capitalising on advancements in deep learning within micro-nano photonics, this paper provides a systematic review of four key methodologies: adjoint method, discriminative models, generative models and reinforcement learning. It covers both inverse design and forward prediction as well as their extended applications in intelligent optical systems. The adjoint method leverages Lorentz reciprocity to overcome the limitation tying gradient computation to parameter dimensionality, requiring only two electromagnetic simulations to obtain high-dimensional parameter gradients. Discriminative models such as multilayer perceptrons (MLPs) and convolutional neural networks (CNNs) establish direct mappings between structural parameters and optical responses, enabling millisecond-level predictions and inverse parameter retrieval for simple devices. Generative models, including variational autoencoders (VAEs) and conditional generative adversarial Networks (cGANs), learn intrinsic distributions of "structure–response" data, addressing the "one-to-many" mapping problem in inverse design and offering diverse topological solutions. Reinforcement learning operates through dynamic "agent-environment" interaction, suiting complex design contexts involving fabrication constraints and multi-objective optimisation. This paper examines current technical challenges, including the adjoint method’s sensitivity to initial designs, the high demand for high-quality training data in generative modelling, and the computational burden of reinforcement learning due to extensive iterations. Finally, it outlines future research directions, such as the integration of physics-driven and data-driven strategies and the design of dynamically reconfigurable devices, aiming to provide insight into the comprehensive application of deep learning techniques in micro-nano photonic device development.