2026/07/01 by Mohammed Abduljaleel Najm, Shaymaa R. Tahhan, Md Mamun Ali +3
Engineering · Physics and Astronomy · #Photonic Crystal and Fiber Optics #Terahertz technology and applications #Photonic Crystals and Applications
paper · doi:10.1016/j.ijleo.2026.172842
This paper describes a deep learning (DL) based framework for accurately and efficiently predicting important electromagnetic parameters in Kagome photonic crystal fibers (PCFs) in the terahertz (THz) regime. A tailored multilayer perceptron (MLP) was trained on high-fidelity FEM simulation data for the prediction of effective refractive indices (neff-X, neff-Y), effective area Aeff, and dispersion with high accuracy. The model achieved R² values higher than 0.9996 and provided a computational speed-up of more than 400,000 times compared to simulation using conventional methods. In contrast to the problem of space complexity, our model needed just 2,464 samples for training and continued to exhibit excellent generalizing performance. The framework proposed here can support inverse design, enabling the retrieval of structural parameters with an error of ≤2.5% in real-time. Optimized designs exhibited ultra-low Aeff (2.76 × 10⁴ μm²) and flatband dispersion flattening (±0.00325 ps/nm·km, 0.8–1.5 THz), surpassing those previously published designs. This work establishes a new benchmark for data-driven PCF design and presents a vision for a transformative impact on terahertz photonics through deep learning.