Machine learning Model Design for two-dimensional Photonic crystals with point defects

Abstract

 The design and optimization of two-dimensional photonic crystals (PhCs) with point defects play a crucial role in advancing nanophotonic devices for applications in sensing, communications, and quantum technologies. This study presents a comprehensive machine learning (ML) framework for predicting and analyzing the optical properties of PhCs with point defects, combining simulation-generated datasets with advanced neural network architectures. Nine tables summarizing structural parameters, defect configurations, and resulting photonic bandgap properties illustrate the quantitative performance of various ML models, while twelve complex figures—including line, bar, pie, scatter, and hybrid plots—visualize the relationships between defect parameters and photonic responses. The results indicate that convolutional neural networks (CNNs) outperform multilayer perceptrons (MLPs) in capturing the nonlinear dependence of resonance frequencies and quality factors on defect geometry. Feature importance analyses further reveal that defect radius, refractive index contrast, and lattice spacing are the most influential parameters, providing insights into design sensitivity. The study also demonstrates that hybrid interpretability techniques can guide rational photonic crystal design while maintaining physical plausibility. Overall, the proposed ML framework reduces computational costs compared to conventional finite-difference time-domain (FDTD) simulations, accelerates design cycles, and enables multi-objective optimization for complex PhC structures. These findings underscore the transformative potential of machine learning in nanophotonics and highlight opportunities for integrating data-driven methods into future photonic device engineering.