Physics-informed recurrent neural network for time dynamics in optical resonances

Resonance structures and features are ubiquitous in optical science. However, capturing their time dynamics in real-world scenarios suffers from long data acquisition time and low analysis accuracy due to slow convergence and limited time windows. Here we report a physics-informed recurrent neural network to forecast the time-domain response of optical resonances and infer corresponding resonance frequencies by acquiring a fraction of the sequence as input. The model is trained in a two-step multi-fidelity framework for high-accuracy forecast, using first a large amount of low-fidelity physical-model-generated synthetic data and then a small set of high-fidelity application-specific data. Through simulations and experiments, we demonstrate that the model is applicable to a wide range of resonances, including dielectric metasurfaces, graphene plasmonics and ultra-strongly coupled Landau polaritons, where our model captures small signal features and learns physical quantities. The demonstrated machine-learning algorithm can help to accelerate the exploration of physical phenomena and device design under resonance-enhanced light–matter interaction. Cascaded gated-recurrent-unit networks trained through a physics-informed multi-fidelity approach can accurately forecast long time sequences and capture their dynamics in a wide range of optical resonance structures and features.