Inverse design epsilon-near-zero-based broadband nonreciprocal thermal emitter using hybrid deep learning framework

Abstract

Nonreciprocal thermal radiation, which breaks Kirchhoff's law by decoupling absorptivity and emissivity, is essential for advanced radiative heat transfer control. However, achieving broadband and tunable nonreciprocal thermal radiation with high design efficiency remains a challenge. This study proposes a novel hybrid deep learning framework, integrating the Artificial Rabbit Optimization and tandem neural network, to inversely design multilayer films (MLFs) based on magnetized gradient epsilon-near-zero (ENZ) InAs layers. By using the Artificial Rabbit Optimization algorithm, we collect a high-quality dataset with a noise ratio of only 3.2%, significantly reducing computational overhead compared to random sampling. The multitasking tandem neural network converges to a low cost function of 0.086, improving the accuracy and avoiding the scattering problem faced by traditional neural networks. Results show that significant nonreciprocal thermal radiation (nonreciprocity > 0.637 and peak 0.723) is achieved in the 14–19 μm range by exploiting the magneto-optical effects of InAs and ENZ-induced Brewster modes. Furthermore, the MLFs exhibit reversed absorptivity and emissivity spectra under reversed magnetic fields. These findings provide a data-efficient and scalable solution for dynamic thermal management and infrared camouflage, demonstrating the powerful synergy between deep learning and nonreciprocal photonics.