Enhancing Nanophotonic Device Inverse Design Through a Class Conditional Generative Adversarial Network with Integrated Classifier on StyleGAN2-ADA Framework

This paper proposes a novel Class Conditional Generative Adversarial Network (CCGAN) tailored for the inverse design of high-resolution nanophotonic device images, with a focus on achieving optimal transmittance at specific wavelengths. The motivation of this research is rooted in the limitations of conventional computational methodologies in nanophotonics, which rely heavily on expert intuition and computational simulators to validate designs. These approaches not only demand extensive computational resources but also bear the risk of escalated costs due to potential design failures. Addressing these challenges, our study leverages the inverse design paradigm, utilizing generative models, to directly derive device structures from desired features. Our proposed CCGAN model introduces significant enhancements to the StyleGAN2-ADA architecture, integrating a classifier that utilizes contrastive loss alongside classification loss to refine the design generation process. This integration facilitates the learning of complex data distributions and the generation of detailed patterns, overcoming the limitations of previous models. The experimental framework employs datasets generated with MaxwellFDFD, a solver for finite-difference frequency-domain (FDFD) Maxwell's equations, categorized by maximum transmittance at specific wavelengths. Quantitative evaluation through the Fréchet Inception Distance (FID) metric demonstrates our model's superior performance, achieving significant reductions in FID values compared to the baseline StyleGAN2-ADA model and other prior models. Moreover, our model also demonstrates superior accuracy in comparison to other models. This research not only advances the field of nanophotonics by providing a robust computational framework for the inverse design of nanodevices but also opens avenues for future exploration into advanced techniques such as diffusion models for further enhancing design accuracy and quality.