Deep learning in PT-symmetric multimode waveguide sensors

We examine the dynamic response of a waveguide with a PT-symmetric complex potential to perturbations in its refractive index. The output transverse intensity profile is recorded as different index perturbations are imposed. The waveguide exhibits its highest sensitivity when it is operating near its exceptional point (EP). A similar behavior is observed when multiple waveguides are coupled and operated around the new EP. A neural network is deployed to decode the cross-sectional intensity images and classify the strengths and positions of the applied perturbations. The proposed waveguide’s behavior near the EP allows the neural network to characterize the perturbations with high accuracy despite noise augmentation, contrary to the Hermitian case. Beyond 1D profiles, this scheme can be readily extended to recover full 2D perturbation distributions. A 1D PT-symmetric lattice structure comprised of five coupled waveguides is able to train a deep neural network capable of reconstructing a 2D perturbation map within the lattice from noisy intensity data, contrary to the Hermitian system. The results show how PT symmetry can be utilized in waveguides to create efficient sensors and imaging devices.