Machine Learning Assisted Broadband Dielectric Photonic Media

Ulusal Optik, Elektro-Optik ve Fotonik Çalıştayı, İstanbul, Türkiye, 12 Eylül 2025, ss.38, (Özet Bildiri)

In this study, an effective negative index photonic medium to enable negative refraction over a broadband frequency

range was designed through the assistance of artificial intelligence. Both frequency-domain and time-domain simulation

packages were employed [1, 2]. In the first step, a dataset was constructed by generating completely random binary-

pattern and all-dielectric photonic crystal geometries and obtaining their corresponding dispersion diagrams () (please

see Fig. 1a). From this dataset, a subset of bands exhibiting effective negative refractive index characteristics was

selected. For each geometry, the negative refractive index value, the degree of isolation of the corresponding band from

other bands, and the operational frequency bandwidth were listed as outputs. In this way, the required input (binary

pattern) and output data for machine learning were prepared. In the next stage, negative refraction of an inclined plane

wave through the photonic medium with a negative effective refractive index was investigated in the time domain

(please see Fig. 1b). While a conventional metamaterial designed with metallic unit cells typically operates only within

a very narrow frequency range (resonant behavior), the machine-learning-based approach adopted in this study enables

the operational frequency bandwidth to be significantly broadened (> 15%). In this way, it would be possible to design a

lens composed entirely of dielectric unit cells such that a broadband electromagnetic field waist at the focal region can

be confined within subwavelength values.