Inverse design and back-substitution validation of dual-EIT photonic crystal structures based on CVAE

Abstract

The dual electromagnetically induced transparency (Dual-EIT) effect in photonic crystals exhibits significant potential for applications in multi-channel sensing and slow-light devices. However, conventional forward design based on the finitedifference time-domain (FDTD) method relies heavily on time-consuming parameter sweeps. Meanwhile, complex micro/nano-optical systems inherently suffer from a severe one-to-many mapping problem (i.e., different structural parameters can produce highly similar spectra), which often causes traditional deep-learning-based inverse design methods to generate physically invalid structures due to parameter averaging. To overcome this limitation, this paper proposes a generative inverse design framework based on a conditional variational autoencoder (CVAE). A total of 5,000 high-fidelity samples were generated using FDTD simulations, among which 2,145 high-quality Dual-EIT positive samples were selected through a rigorous peak-detection algorithm. Conditioned on the target transmission spectrum, the CVAE successfully maps high-dimensional structural parameters into a latent probabilistic space. The experimental results show that the mean absolute error (MAE) of the predicted geometric parameters is strictly controlled within 0.3 nm. More importantly, a complete closed-loop validation framework of “prediction–physical back-substitution” is established. By re-inserting the predicted parameters into FDTD simulations, the results confirm that the model captures the underlying physical coupling mechanisms rather than merely performing numerical fitting, effectively overcoming the validation challenges caused by the extreme nonlinear sensitivity of high-Q dark modes.