Strategical Deep Learning for Photonic Bound States in the Continuum

Resonance is instrumental in modern optics and photonics. While one can use numerical simulations to sweep geometric and material parameters of optical structures, these simulations usually require considerably long calculation time and substantial computational resources. Such requirements significantly limit their applicability in the inverse design of structures with desired resonances. The recent introduction of artificial intelligence allows for faster spectra predictions of resonance. However, even with relatively large training datasets, current end‐to‐end deep learning approaches generally fail to predict resonances with high‐quality‐factors (Q‐factor) due to their intrinsic non‐linearity and complexity. Here, a resonance informed deep learning (RIDL) strategy for rapid and accurate prediction of the optical response for ultra‐high‐Q‐factor resonances is introduced. By incorporating the resonance information into the deep learning algorithm, the RIDL strategy achieves a high‐accuracy prediction of reflection spectra and photonic band structures while using a comparatively small training dataset. Further, the RIDL strategy to develop an inverse design algorithm for designing a bound state in the continuum (BIC) with infinite Q‐factor is applied. The predicted and measured angle‐resolved band structures of this device show minimal differences. The RIDL strategy is expected to be applied to many other physical phenomena such as Gaussian and Lorentzian resonances. Photonic bound states in the continuum (BICs) are instrumental for many applications, but designing them is extremely resource‐consuming due to their high‐quality nature. With the help of the adaptive data acquisition (ADA) method, which reduced ∼99% of simulation workload, the resonance‐informed deep learning (RIDL) strategy significantly increased the prediction accuracy and reduced computational time for designing BICs with ultra‐high‐quality‐factors.