Deep learning design of nanoscale polariton propagations in twisted van der Waals multilayers

Nano-optics aims to understand and control the propagation of light at the nanoscale through the excitation of surface polaritons: hybrid light-matter quasiparticles. Recently, twisted van der Waals materials have enabled unprecedented phonon polariton propagations, such as canalization. However, nano-optics still presents an important limitation: obtaining polariton propagations on demand. Here we combine deep neural networks with twisted polaritonic multilayers to enable on-demand design of phonon polariton propagation. We demonstrate canalization, bicanalization and tricanalization in twisted α-MoO₃ homostructures over previously unexplored frequencies (600-800 cm^-1). We illustrate the practical potential of our method by achieving a desired polariton propagation in an existing α-MoO₃ bilayer by adding an extra α-MoO₃ layer. Finally, we extend our neural networks to a variety of other materials, allowing us to predict canalization from the visible to the terahertz regime. Our deep-learning-based approach offers considerable potential for advancing nanophotonic applications in areas such as sensing or thermal management.